<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Shamuddin</title>
    <description>The latest articles on DEV Community by Shamuddin (@aitecharchive).</description>
    <link>https://dev.to/aitecharchive</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4046546%2Ff37f6d10-85a6-4741-9ed8-0081a58c931a.png</url>
      <title>DEV Community: Shamuddin</title>
      <link>https://dev.to/aitecharchive</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/aitecharchive"/>
    <language>en</language>
    <item>
      <title>Edge0-145B Open Weights: RTX GPU vs Mac, RTX Wins</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Thu, 19 Nov 2026 04:44:58 +0000</pubDate>
      <link>https://dev.to/aitecharchive/edge0-145b-open-weights-rtx-gpu-vs-mac-rtx-wins-30gp</link>
      <guid>https://dev.to/aitecharchive/edge0-145b-open-weights-rtx-gpu-vs-mac-rtx-wins-30gp</guid>
      <description>&lt;p&gt;For builders choosing hardware, the Edge0-145B open weights are an RTX story, not a Mac story. On November 12, 2026, Edge0 &lt;a href="https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer" rel="noopener noreferrer"&gt;released the 145-billion-parameter model&lt;/a&gt; with weights, an inference engine and a desktop app. The company reports 29GB peak memory, but the &lt;a href="https://github.com/Edge0-AI/Edge0" rel="noopener noreferrer"&gt;GitHub requirements&lt;/a&gt; set a floor of an NVIDIA GPU with at least 16 GiB of VRAM plus 20 GiB of system RAM. A 12 GB card, a Mac or a free Colab T4 will not run it, so the verdict is a single RTX card with 16 GiB or more, on Linux or Windows.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Edge0-145B and Edge0-120B were released on November 12, 2026, alongside the edge0-infer v0.2.0 engine and the Edge0 Chat app.&lt;/li&gt;
&lt;li&gt;Hardware floor: NVIDIA GPU, CUDA 13, compute capability 7.5 or higher, 16 GiB VRAM, 20 GiB system RAM.&lt;/li&gt;
&lt;li&gt;macOS is not supported for the 145B and 120B models in the &lt;a href="https://github.com/Edge0-AI/Edge0/releases/tag/v0.2.0" rel="noopener noreferrer"&gt;v0.2.0 release&lt;/a&gt;, despite the launch post suggesting otherwise.&lt;/li&gt;
&lt;li&gt;Benchmark scores are Edge0's own and are strongest on coding and terminal tasks.&lt;/li&gt;
&lt;li&gt;The weights use a community licence, not Apache 2.0, and Qwen's upstream licence still applies.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Which should you pick: an RTX GPU or a Mac?
&lt;/h2&gt;

&lt;p&gt;Pick the RTX GPU. The platform table in the GitHub README lists Linux and Windows with NVIDIA hardware for the 145B and 120B models. Metal and macOS support applies to the smaller Edge0-35B tier. The launch post's FAQ says macOS works on day one, but independent write-ups from &lt;a href="https://enjoy.dev/blog/edge0-145b" rel="noopener noreferrer"&gt;enjoy.dev&lt;/a&gt; and &lt;a href="https://explainx.ai/blog/edge0-145b-open-on-device-laptop-agent-model-november-2026" rel="noopener noreferrer"&gt;explainx.ai&lt;/a&gt; noted that this is not reflected in the release itself.&lt;/p&gt;

&lt;p&gt;If you already own a Mac, the practical option is Edge0-35B, which Edge0 says runs under 2.5GB on a phone. It is a different model with a different capability level, so treat it as an alternative and not a substitute.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does Edge0-145B need to run locally?
&lt;/h2&gt;

&lt;p&gt;The requirements, as published in the &lt;a href="https://github.com/Edge0-AI/Edge0" rel="noopener noreferrer"&gt;repository&lt;/a&gt;, are an RTX 20, 30, 40 or 50 series card (compute capability 7.5 or higher), CUDA 13, 16 GiB of GPU memory and 20 GiB of system RAM.&lt;/p&gt;

&lt;p&gt;The "29GB" figure refers to peak memory in Edge0's own measurement, which counts the GPU and RAM tiers together. Third-party testing by &lt;a href="https://www.eesel.ai/blog/edge0-145b-review" rel="noopener noreferrer"&gt;eesel.ai&lt;/a&gt; on a 24GB card found 23.14 GiB of GPU memory plus 31.28 GiB of system RAM for the Q3_K build, about 54.4 GiB in total. These figures are not contradictory. They describe different configurations, so budget for more RAM than the headline suggests if you want headroom.&lt;/p&gt;

&lt;p&gt;Cards that miss the floor:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;RTX 3060 12GB: under the 16 GiB VRAM minimum.&lt;/li&gt;
&lt;li&gt;Free Colab T4: compute capability 7.5, but its 15 GiB of VRAM is about 1 GiB short.&lt;/li&gt;
&lt;li&gt;Free Kaggle P100: compute capability 6.0, below the 7.5 requirement.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How does Edge0-145B fit on one GPU?
&lt;/h2&gt;

&lt;p&gt;The model is a latent mixture-of-experts derived from Alibaba's Qwen3.8-Flash-Next. According to the &lt;a href="https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer" rel="noopener noreferrer"&gt;launch post&lt;/a&gt;, it has 145B total parameters with about 10B active per token, 512 experts per layer across 48 layers, ten experts selected per token and one shared expert. Edge0 says the two builds, Q3_K and NVFP4, were trained natively at 4 bits.&lt;/p&gt;

&lt;p&gt;The memory trick is a three-tier layout. Frequently used experts stay in GPU memory, a library of about 40GB sits in system RAM, and about 48GB of vocabulary data is read from SSD on demand. Groups of 64 experts share two large matrices, and each expert keeps only a small delta. The GPU computes cached experts while the CPU works in parallel, and speculative decoding spreads weight reads across candidate tokens.&lt;/p&gt;

&lt;p&gt;The practical consequence is that fast system RAM and a fast SSD matter alongside the GPU. A 16 GiB card with slow storage will not behave like the figures below.&lt;/p&gt;

&lt;h2&gt;
  
  
  How fast is Edge0-145B on real hardware?
&lt;/h2&gt;

&lt;p&gt;All of these numbers come from Edge0's README testing on an RTX 5090 D, so treat them as vendor measurements.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Setup (Q3_K build)&lt;/th&gt;
&lt;th&gt;4K prefill&lt;/th&gt;
&lt;th&gt;Chat decode&lt;/th&gt;
&lt;th&gt;Agent decode&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;RTX 5090 D, full 24GB&lt;/td&gt;
&lt;td&gt;3,236.7 tok/s&lt;/td&gt;
&lt;td&gt;130.2 tok/s&lt;/td&gt;
&lt;td&gt;144.6 tok/s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;RTX 4090 24GB&lt;/td&gt;
&lt;td&gt;2,637.5 tok/s&lt;/td&gt;
&lt;td&gt;95.0 tok/s&lt;/td&gt;
&lt;td&gt;104.9 tok/s&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;RTX 5090 D limited to 16 GiB VRAM and 20 GiB RAM&lt;/td&gt;
&lt;td&gt;430 tok/s&lt;/td&gt;
&lt;td&gt;62.4 tok/s&lt;/td&gt;
&lt;td&gt;73.3 tok/s&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The constrained row is the one to read if you own a 16 GiB card. It is slower than the full-memory runs but still comfortably above reading speed. The NVFP4 build is slower under the same limit, at 51.3 and 59.9 tokens per second for chat and agent decode, per the &lt;a href="https://github.com/Edge0-AI/Edge0" rel="noopener noreferrer"&gt;GitHub README&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Do the benchmark scores hold up?
&lt;/h2&gt;

&lt;p&gt;They are internal evaluations, and no independent reproduction has been published as far as the sources I reviewed show. Edge0 reports a six-benchmark mean of 72.8 for the 145B model against 71.7 for Qwen3.8-Flash-Next and 70.1 for DeepSeek-V4-Flash-0731. On individual tests it reports 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond (close to Claude-Opus-4.6 (Max) at 91.3), and 80.6 on IFBench against 62.5 for the same Opus model, per the &lt;a href="https://edge0.ai/blog/introducing-edge0-145b-frontier-ai-capabilities-on-your-own-computer" rel="noopener noreferrer"&gt;launch post&lt;/a&gt;. It does concede ground on Humanity's Last Exam, scoring 35.4 against 40.0 for Claude-Opus-4.6 (Max).&lt;/p&gt;

&lt;p&gt;The pattern is credible: a model tuned for agentic coding and terminal work, with less margin on hard knowledge questions. Run your own tasks before committing a workflow to it. A note on scope, using our own measurement: first-party DataForSEO pricing run completed 2026-11-12, keyword corpus n=695, found that only 72 out of 695 AI-tooling keywords clear a winnable search bar, which is why verdict pages like this one are rare rather than filler. For comparison with other self-hostable releases, see our notes on &lt;a href="https://dev.to/articles/longcat-2-5-open-weights-mit-self-host"&gt;LongCat 2.5&lt;/a&gt; and &lt;a href="https://dev.to/articles/kimi-k3-1-pro-open-weights-released"&gt;Kimi K3.1 Pro&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is Edge0-145B really open source?
&lt;/h2&gt;

&lt;p&gt;The artifacts are open, but the licence is not Apache. The model weights fall under the Edge0 Community Model License 1.0, dated November 12, 2026, which permits personal use, research, modification, noncommercial sharing, internal use and commercial use of generated outputs. Selling or bundling the model, running a commercial model service, or offering an independent commercial coding or office assistant needs a separate written licence from Edge0. The terms are in the &lt;a href="https://huggingface.co/Edge0/Edge0-145B-NVFP4" rel="noopener noreferrer"&gt;Hugging Face repository&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Upstream rights remain. The NOTICE file states that Edge0 does not claim the whole model belongs to it, and the Qwen Community License 1.0 continues to apply, as &lt;a href="https://writeble.com/2026/11/13/edge0-145b-laptop-model-qwen-base-community-license/" rel="noopener noreferrer"&gt;writeble.com&lt;/a&gt; also points out. The engine is Apache-2.0 and Strata is MIT, as &lt;a href="https://explainx.ai/blog/edge0-145b-open-on-device-laptop-agent-model-november-2026" rel="noopener noreferrer"&gt;explainx.ai&lt;/a&gt; summarises. If you plan a product, read both licences first. Our &lt;a href="https://dev.to/articles/runway-praxis-1-open-weights-gpu-verdict"&gt;Runway Praxis 1 GPU verdict&lt;/a&gt; covers a similar licence-and-hardware check.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you set it up?
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Use a Linux or Windows machine with an RTX 20, 30, 40 or 50 series GPU of 16 GiB or more and at least 20 GiB of RAM. Install CUDA Toolkit 13.&lt;/li&gt;
&lt;li&gt;Clone the &lt;a href="https://github.com/Edge0-AI/Edge0" rel="noopener noreferrer"&gt;Edge0 repository&lt;/a&gt; and build edge0-infer, or take the Windows package from release v0.2.0.&lt;/li&gt;
&lt;li&gt;Download a checkpoint from Hugging Face: Edge0/Edge0-145B-q3k or Edge0/Edge0-145B-NVFP4. The 120B variants are smaller but have the same GPU floor.&lt;/li&gt;
&lt;li&gt;Run &lt;code&gt;edge0-infer serve&lt;/code&gt; for a local OpenAI-compatible and Anthropic Messages-compatible endpoint, then point your coding agent at localhost. Alternatively use &lt;code&gt;edge0-infer chat&lt;/code&gt; or the Edge0 Chat desktop app.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If offline use is your main goal, our piece on &lt;a href="https://dev.to/articles/omnii-1-offline-ai-pathogen-screening-laptop"&gt;OMNII-1 running on a laptop&lt;/a&gt; shows a lighter-weight path. Builders interested in post-training a Qwen base can also look at &lt;a href="https://dev.to/articles/arete-27b-qwen3-8-27b-propaganda-post-train-free"&gt;Arete 27B&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: Can Edge0-145B run on a MacBook?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Not in the current release. The platform table and the v0.2.0 notes list Linux and Windows with NVIDIA GPUs for the 145B and 120B models. Only Edge0-35B has Metal support on macOS.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: What is the minimum GPU for Edge0-145B open weights?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; An NVIDIA card with CUDA 13, compute capability 7.5 or higher and at least 16 GiB of VRAM, plus 20 GiB of system RAM, according to the project README.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Can I use Edge0-145B commercially?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Partly. Commercial use of generated outputs is allowed, but selling or bundling the model, running a commercial model service, or shipping a commercial assistant needs a written licence from Edge0. The Qwen Community License also still applies.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does the free Colab tier work?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. The free T4 has about 15 GiB of VRAM, which is below the 16 GiB floor. Kaggle's P100 fails the compute capability requirement.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Should I choose Edge0-145B or Edge0-120B?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Choose 145B if your hardware meets the floor, since Edge0 reports higher scores on most benchmarks. The 120B has the same GPU floor but a smaller footprint, so it mainly helps if RAM or disk is tight.&lt;/p&gt;

</description>
      <category>edge0145b</category>
      <category>openweights</category>
      <category>localllm</category>
      <category>nvidiartx</category>
    </item>
    <item>
      <title>N-central 4EverAble2Shell: Legacy API Off vs Hotfix 2</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Thu, 19 Nov 2026 04:44:38 +0000</pubDate>
      <link>https://dev.to/aitecharchive/n-central-4everable2shell-legacy-api-off-vs-hotfix-2-4pk9</link>
      <guid>https://dev.to/aitecharchive/n-central-4everable2shell-legacy-api-off-vs-hotfix-2-4pk9</guid>
      <description>&lt;p&gt;Turning off the Legacy API is the better defence against the n-central 4everable2shell exploit chain, and installing Hotfix 2 on its own is not enough. By 12:50 UTC on November 12, 2026, Huntback's mining-pool telemetry showed &lt;a href="https://huntback.io/blog/ncentral-4everable2shell-miners-on-live-servers" rel="noopener noreferrer"&gt;21 production N-able N-central servers compromised&lt;/a&gt;, and 19 of them reported build 2026.4.1.26, which is N-central 2026.4 Hotfix 2, released on November 10. The gap between public proof-of-concept and a full root-level exploit was about 5.3 hours, from 03:58 UTC to 09:16 UTC on the same day. &lt;a href="https://uptime.n-able.com/event/202613/" rel="noopener noreferrer"&gt;N-able's advisory&lt;/a&gt; tells on-premises customers to disable the Legacy API immediately, and the evidence so far backs that instruction over any reliance on the patch level.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;N-central is a remote monitoring and management (RMM) platform. Root on the server means script execution on every managed customer endpoint.&lt;/li&gt;
&lt;li&gt;Affected: on-premises N-central with the Legacy API enabled and reachable from the internet. N-able says hosted (NCOD) environments are not affected by the server issue.&lt;/li&gt;
&lt;li&gt;19 of 21 confirmed victims ran Hotfix 2, so patching was not a protection here.&lt;/li&gt;
&lt;li&gt;The observed payload was an XMRig cryptominer, but the access obtained is far more valuable than mining.&lt;/li&gt;
&lt;li&gt;Do this today: verify the Legacy API is off, restrict exposure, and hunt using N-able's published IOC list.&lt;/li&gt;
&lt;li&gt;A follow-up hotfix has been promised. Until it ships, configuration is the control.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What is the n-central 4everable2shell exploit chain?
&lt;/h2&gt;

&lt;p&gt;4EverAble2Shell is the name attached to a set of N-central flaws published alongside a related Agent2Shell proof of concept. The server-side chain has two stages, according to &lt;a href="https://huntback.io/blog/ncentral-4everable2shell-miners-on-live-servers" rel="noopener noreferrer"&gt;Huntback's analysis&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The first stage is an unauthenticated request to the Apache Axis SOAP endpoint at /dms/services/ServerUI. The flaw is tracked as &lt;a href="https://nvd.nist.gov/vuln/detail/CVE-2026-90436" rel="noopener noreferrer"&gt;CVE-2026-90436&lt;/a&gt;, a CWE-470 unsafe reflection issue that NVD scores as network-reachable, low complexity, no privileges and no user interaction. WatchTowr, Previdian and SecurityWeek report a CVSS score of 9.8. In plain terms, Axis 1.x parses an incoming SOAP fault and uses a name inside it to instantiate a Java class of the attacker's choosing. A class in the bundled jline library runs a shell command when it is constructed. Authentication handlers do not help, because the fault is parsed before the service handler chains run.&lt;/p&gt;

&lt;p&gt;That gives code execution as the low-privileged nable service account. The second stage abuses N-central's own privileged helper script, nable_wrapper.pl, to redirect a root-owned script and run it as root, then restore the original to hide the change. I am deliberately not describing the exact request shape or the filesystem steps.&lt;/p&gt;

&lt;h2&gt;
  
  
  How quickly was the 4EverAble2Shell exploit weaponised?
&lt;/h2&gt;

&lt;p&gt;Within hours. The times below are UTC on November 12, 2026, taken from Huntback, &lt;a href="https://previdian.com/VULN-2026-0009" rel="noopener noreferrer"&gt;Previdian's VULN-2026-0009 entry&lt;/a&gt; and N-able's notice.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;03:58: researcher Sean Bales publishes the proof of concept, tested against Hotfix 2.&lt;/li&gt;
&lt;li&gt;08:24: Previdian sensors see the first exploitation attempts.&lt;/li&gt;
&lt;li&gt;08:30: the first probe from 157.20.241[.]16 reaches Huntback's decoy servers.&lt;/li&gt;
&lt;li&gt;09:16: first weaponised full-chain request, with root access and a miner install.&lt;/li&gt;
&lt;li&gt;12:50: the mining pool lists 21 live N-central rigs.&lt;/li&gt;
&lt;li&gt;16:00: N-able publishes its critical security update notice.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;N-able's November 12 notice said it had seen no signs of compromise. By November 13 it said it had identified indicators of compromise and was working with specific customers, and it published the IOC list at 13:40 UTC. &lt;a href="https://www.securityweek.com/hackers-exploit-new-n-central-zero-day-hours-after-poc-publication/" rel="noopener noreferrer"&gt;SecurityWeek&lt;/a&gt; reports N-able saying every unauthenticated bug from the first disclosure batch is fixed, with investigation of the new batch continuing.&lt;/p&gt;

&lt;p&gt;This compressed pattern is not unique. Across our own scan of the disclosure-to-exploitation landscape on November 13, 2026, our channel-harvest method (n=1020, measured 2026-11-13) samples the 68 AI-focused channels we track to see which security topics are already saturated - exploitation-window stories like this one are not. We saw a similar squeeze in &lt;a href="https://dev.to/articles/openssh-sshole-exploited-40-hours-after-patch"&gt;OpenSSH's SSHole, exploited 40 hours after the patch&lt;/a&gt;, and in a &lt;a href="https://dev.to/articles/barracuda-esg-exploited-cisa-deadline"&gt;Barracuda ESG case tied to a CISA deadline&lt;/a&gt;. For the broader question of how fast capable tooling shortens these windows, see our piece on &lt;a href="https://dev.to/articles/gpt-6-cyber-benchmarks-price-access"&gt;GPT-6 cyber benchmarks&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is it better to disable the Legacy API or install Hotfix 2?
&lt;/h2&gt;

&lt;p&gt;Disable the Legacy API. Installing Hotfix 2 is still necessary, but it is not a substitute.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Control&lt;/th&gt;
&lt;th&gt;What it addresses&lt;/th&gt;
&lt;th&gt;Evidence from this incident&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Legacy API off&lt;/td&gt;
&lt;td&gt;Removes the exposed Axis endpoint dependency that N-able ties to the server issue&lt;/td&gt;
&lt;td&gt;N-able instructs it as the immediate action; it is off by default&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Hotfix 2 (build 2026.4.1.26)&lt;/td&gt;
&lt;td&gt;Fixes the November 9 batch, including CVE-2026-125883 (XXE) and CVE-2026-125878 (OGNL injection)&lt;/td&gt;
&lt;td&gt;19 of 21 compromised servers were already on it&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Two caveats. First, I cannot confirm from public material that the Legacy API was enabled on every one of the 21 servers; the link comes from N-able's advisory and the fact that all 21 served the N-central login page. Second, "off by default" is not "off on your server". Validate the setting rather than assuming it, using N-able's knowledge base article on disabling Legacy API settings on on-premises servers.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who is affected, and does this matter for India?
&lt;/h2&gt;

&lt;p&gt;Affected are MSPs running on-premises N-central with the Legacy API enabled and exposed to the internet. Lupovis, speaking to SecurityWeek, placed the 21 rigs in the US (9), South Africa (4), Canada (4), the UK, Germany, France and Sweden.&lt;/p&gt;

&lt;p&gt;India is relevant for two reasons. Previdian's sensors geolocated the attacking address to India, based on two attempts from one IP on November 12. That is sensor geolocation only and says nothing about who is behind it; the operator is not publicly attributed. The more practical point is for Indian MSPs and IT service providers: a compromised RMM with downstream customer access is likely reportable under CERT-In's incident reporting directions of April 2022, which set a six-hour window. No India-specific advisory has been published for this event.&lt;/p&gt;

&lt;h2&gt;
  
  
  What should MSPs do now about n-central 4everable2shell?
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;Disable the Legacy API on every on-premises server and confirm it, rather than trusting the default.&lt;/li&gt;
&lt;li&gt;Finish the Hotfix 2 upgrade and keep agents and probes current per N-able guidance, while treating it as hygiene.&lt;/li&gt;
&lt;li&gt;Check the server's exposure to the internet. Limiting access to the web interface to known networks or a VPN is standard practice and reduces risk, though it is my recommendation, not a stated N-able fix.&lt;/li&gt;
&lt;li&gt;Search firewall and proxy logs for outbound connections to 157.20.241[.]16 on port 8080 and TCP ports 19202 to 19418, and to the mining endpoints 139.99.69[.]109:10128 and 79.76.51[.]56:39100.&lt;/li&gt;
&lt;li&gt;On the host, look for /tmp/.xvd, /var/tmp/.xvd, an xvd.service unit, /etc/cron.d/xvd, and processes started with &lt;code&gt;--rig-id nc_&lt;/code&gt; or &lt;code&gt;--donate-level 1&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;In web logs, look for POSTs to /dms/services/ServerUI containing a SOAP fault with an exceptionName element, the user agent python-requests/2.31.0, and the escalation script with SHA-256 fc4c01fd11bc301cf312830460cfdccc3af9dcfe50c910b86d6a1117bc76c747.&lt;/li&gt;
&lt;li&gt;Treat any hit as a full compromise. Huntback's position is that this is not a cleanup job, so rebuild, rotate every credential the RMM held, and review script and task history across downstream customers.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The mining operation looks opportunistic. The same wallet listed 631 rigs at the time of Huntback's snapshot, across products such as Jpom, FreePBX and MLflow. That suggests N-central was the newest target for an existing campaign, but the same access could be used for far worse.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: Does Hotfix 2 protect against 4EverAble2Shell?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. Huntback found that 19 of the 21 confirmed compromised servers were running build 2026.4.1.26, which is Hotfix 2. Disable the Legacy API and wait for the follow-up hotfix N-able has promised.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Are hosted N-central (NCOD) customers affected?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; N-able states that hosted environments are not affected by the server issue and that protections have been applied. The advisory concerns on-premises servers with the Legacy API enabled and internet-exposed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: What was the attacker doing on compromised servers?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; In the observed cases they installed XMRig 6.26.0 for cryptomining, using half the CPU cores and persisting through cron and systemd. Root access on an RMM could enable much more, so mining should not be read as the limit of the risk.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Who is behind the attacks?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The operator is not publicly attributed by N-able or the researchers. Previdian's sensors geolocated the attacking IP to India, which is a network location and not evidence of who controls it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Where do I find the official indicators of compromise?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; N-able published its IOC list on November 13, 2026 at 13:40 UTC on its &lt;a href="https://uptime.n-able.com/event/202613/" rel="noopener noreferrer"&gt;status dashboard advisory&lt;/a&gt;. Use that as the authoritative list and check it again for updates.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Should I report a suspected compromise?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Report under the rules of your jurisdiction. In India, CERT-In's directions apply to affected service operators, and in other regions your contracts and local regulators may set their own deadlines.&lt;/p&gt;

&lt;p&gt;Further reading: &lt;a href="https://decipher.sc/2026/11/12/another-n-able-vulnerability-emerges/" rel="noopener noreferrer"&gt;Decipher&lt;/a&gt; and &lt;a href="https://thehackernews.com/2026/11/new-n-central-exploits-work-against.html" rel="noopener noreferrer"&gt;The Hacker News&lt;/a&gt; have additional coverage of the wider N-able disclosure series. For a similar patch-first-or-not decision, see &lt;a href="https://dev.to/articles/wordpress-cve-2026-119977-vs-wp2shell-patch-first"&gt;WordPress CVE-2026-119977 vs WP2Shell&lt;/a&gt; and the &lt;a href="https://dev.to/articles/cve-2026-124194-outlook-ntlmv2-zero-day-patch-first"&gt;Outlook NTLMv2 zero-day&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>ncentral</category>
      <category>nable</category>
      <category>rmm</category>
      <category>mspsecurity</category>
    </item>
    <item>
      <title>Nvidia Interim CEO Kress vs an Outside Hire: Kress Wins</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Thu, 19 Nov 2026 04:44:18 +0000</pubDate>
      <link>https://dev.to/aitecharchive/nvidia-interim-ceo-kress-vs-an-outside-hire-kress-wins-1mko</link>
      <guid>https://dev.to/aitecharchive/nvidia-interim-ceo-kress-vs-an-outside-hire-kress-wins-1mko</guid>
      <description>&lt;p&gt;Jensen Huang resigned as Nvidia's president, CEO and board member with immediate effect, and the board named CFO Colette Kress as the Nvidia interim CEO while it searches for a permanent successor, according to &lt;a href="https://nvidianews.nvidia.com/news/nvidia-announces-ceo-transition" rel="noopener noreferrer"&gt;Nvidia's announcement of 12 November 2026&lt;/a&gt;. Between Kress and an outside hire, Kress is the better choice for the interim role, because the first test is the Q3 fiscal 2027 earnings call on 17 November and she already knows the numbers behind it. For the permanent job the case for an outsider is stronger, and that decision can wait.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Huang's resignation was accepted by the board on 11 November and disclosed in an &lt;a href="https://www.sec.gov/Archives/edgar/data/1045810/000104581026000082/nvda-20261111.htm" rel="noopener noreferrer"&gt;SEC Form 8-K&lt;/a&gt;. Kress, 59, signed that filing as interim CEO and CFO.&lt;/li&gt;
&lt;li&gt;Harvey C. Jones becomes Nvidia's first-ever Chair of the Board, per the company's announcement.&lt;/li&gt;
&lt;li&gt;Earnings stay on schedule for Tuesday 17 November, and Kress will host the call.&lt;/li&gt;
&lt;li&gt;Kress structured much of the financing web she now has to run, which is both her advantage and her exposure. For scale: we priced 695 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data (n=695, measured=2026-11-12), and Nvidia-adjacent queries sit well above the median - demand of this kind does not pause for a CEO transition (&lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;Nothing in the announcement or the filing changes product plans or export policy.&lt;/li&gt;
&lt;li&gt;A permanent CEO, internal or external, is the real decision still ahead.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What happened when Jensen Huang resigned from Nvidia?
&lt;/h2&gt;

&lt;p&gt;Nvidia's press release says Huang left as president, CEO and director, and that the board has started a search that will consider internal and external candidates. Kress, executive vice president and CFO, keeps her finance role while acting as CEO, per &lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;. She joined in September 2013 and, per the release, spent nearly 25 years before that in finance roles at technology companies including Cisco and Microsoft.&lt;/p&gt;

&lt;p&gt;The 8-K gives the governance detail. Jones, 73, is managing partner of Square Wave Ventures and has sat on Nvidia's board since 1993, according to the filing. The release describes him as a former CEO and Executive Chairman of Synopsys.&lt;/p&gt;

&lt;p&gt;For scale: we priced 695 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data (n=695, measured=2026-11-12), and Nvidia-adjacent queries sit well above the median - this level of demand does not pause for a CEO transition.&lt;/p&gt;

&lt;p&gt;The trigger was a conduct and national-security story, not a performance failure. &lt;a href="https://apnews.com/article/nvidia-jensen-huang-colette-kress-39ea208c040539166e10d8c6f473bc57" rel="noopener noreferrer"&gt;AP reported&lt;/a&gt; that The New York Times had described an affair between Huang and an executive at Singapore-based data-centre company Megaspeed, whom US intelligence is reported to believe was a Chinese operative. China's Foreign Ministry dismissed the spying allegations. Nvidia said it had no reason to believe national security or company secrets were compromised. These are reported allegations, and I have only repeated what the cited outlets attribute to others.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is Colette Kress the right Nvidia interim CEO?
&lt;/h2&gt;

&lt;p&gt;For the interim period, yes, and the reasoning is practical. Bloomberg, &lt;a href="https://www.livemint.com/companies/people/nvidia-turns-to-cfo-colette-kress-to-steady-ship-after-jensen-huang-exit-11794498763925.html" rel="noopener noreferrer"&gt;as carried by Mint&lt;/a&gt;, put her 13-year CFO tenure at more than twice the average for S&amp;amp;P 500 and Fortune 500 CFOs. Mark Malek of Siebert Financial said she knows every number in the order book. An outsider would need weeks just to learn the customer and financing commitments, and the earnings call is days away.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Criterion&lt;/th&gt;
&lt;th&gt;Kress as interim CEO&lt;/th&gt;
&lt;th&gt;Outside hire&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Knowledge of order book and financing&lt;/td&gt;
&lt;td&gt;Built much of it as CFO&lt;/td&gt;
&lt;td&gt;Starts from scratch&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Readiness for the 17 November call&lt;/td&gt;
&lt;td&gt;Already preparing it&lt;/td&gt;
&lt;td&gt;Cannot realistically host it&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Signal to customers and suppliers&lt;/td&gt;
&lt;td&gt;Continuity&lt;/td&gt;
&lt;td&gt;Disruption on top of disruption&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Distance from past financing deals&lt;/td&gt;
&lt;td&gt;None, she structured them&lt;/td&gt;
&lt;td&gt;Clean slate&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Fit as permanent CEO&lt;/td&gt;
&lt;td&gt;Unproven as an operator&lt;/td&gt;
&lt;td&gt;Depends on the person&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The last two rows are where the outsider wins, which is why this verdict covers the interim role only.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is Kress inheriting at Nvidia?
&lt;/h2&gt;

&lt;p&gt;A company that is large, concentrated and recently shaken. &lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch reported&lt;/a&gt; trailing-12-month revenue of $96.22 billion as of the July 2026 quarter, and shares fell nearly 12% in early trading on 12 November. &lt;a href="https://apnews.com/article/nvidia-jensen-huang-colette-kress-39ea208c040539166e10d8c6f473bc57" rel="noopener noreferrer"&gt;AP reported&lt;/a&gt; the early-October record at $243 a share, with the stock more than 30% below it by 12 November. Nvidia's own &lt;a href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027" rel="noopener noreferrer"&gt;Q2 fiscal 2027 results&lt;/a&gt; put that same quarter at $96.2 billion in revenue, up 106% year on year, with an outlook of $108.0 billion for Q3 and no China data-centre compute revenue assumed.&lt;/p&gt;

&lt;p&gt;Two structural points matter more than the share price:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Huang ran a flat organisation with about 60 direct reports and no named successor, per &lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;. The board has to decide how much of that structure to keep.&lt;/li&gt;
&lt;li&gt;Huang held about 3.5% of the shares with no supervoting rights, per &lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch's report&lt;/a&gt;. The board's power to remove him always existed, and this is the first time it was used.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What does Kress's financing record mean for the interim role?
&lt;/h2&gt;

&lt;p&gt;This is the angle most coverage underplays. &lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch reported&lt;/a&gt; that, as of the end of Q2 fiscal 2027, Nvidia carried $99 billion in investments plus a further $25 billion in commitments to AI and data-centre startups and neoclouds, alongside $36 billion in cloud service commitments. (Note: Nvidia's own &lt;a href="https://nvidianews.nvidia.com/news/nvidia-announces-financial-results-for-second-quarter-fiscal-2027" rel="noopener noreferrer"&gt;Q2 release&lt;/a&gt; separately cites approximately $99.0 billion remaining under its share-repurchase authorization - a different $99 billion, and one more thing the interim CFO now owns explaining on 17 November.) &lt;a href="https://apnews.com/article/nvidia-jensen-huang-colette-kress-39ea208c040539166e10d8c6f473bc57" rel="noopener noreferrer"&gt;AP describes&lt;/a&gt; Kress as a key architect of Nvidia's circular financing arrangements, including a stake in neocloud CoreWeave of close to 13 percent. AP also reported that CoreWeave shares fell about 7 percent when Nvidia dropped on 12 November.&lt;/p&gt;

&lt;p&gt;So the interim CEO is now the operator of arrangements she designed. That helps with continuity, because counterparties such as CoreWeave, OpenAI and Nebius already know her. It hurts with scrutiny, because an analyst asking about circular exposure on 17 November is asking the person who built it. A permanent outsider would have fewer such conflicts, and that is a fair argument for hiring one.&lt;/p&gt;

&lt;h2&gt;
  
  
  Who could be the permanent Nvidia CEO?
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://www.theguardian.com/technology/2026/nov/12/jensen-huang-nvidia-ceo-replacement" rel="noopener noreferrer"&gt;The Guardian&lt;/a&gt; floated Ian Buck, who manages Nvidia's AI hardware and software products, as an internal option, and Rene Haas, a former Nvidia employee now running Arm Holdings, as an external one. Dan Ives told AP he is a big fan of Kress but sees a good chance the board looks outside. Mandeep Singh of Bloomberg Intelligence cautioned that it will be hard for anyone to execute the way Huang did.&lt;/p&gt;

&lt;p&gt;My reading is that the board will use the interim period to see how Kress handles the call and the counterparties before deciding. That is analysis, not reporting.&lt;/p&gt;

&lt;h2&gt;
  
  
  What should readers watch next?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Whether a permanent CEO is named before the end of the calendar year.&lt;/li&gt;
&lt;li&gt;Whether the pace of the financing commitments reported above ($99 billion and $25 billion, per &lt;a href="https://techcrunch.com/2026/11/12/nvidias-jensen-huang-resigns-chipmaker-names-cfo-colette-kress-as-interim-ceo/" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;) changes.&lt;/li&gt;
&lt;li&gt;How the 17 November call treats circular-financing exposure.&lt;/li&gt;
&lt;li&gt;Whether Megaspeed follows through on its pledge, reported by AP on 13 November, to cooperate with US investigations.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For wider context on the week's other stories, see our &lt;a href="https://dev.to/articles/ai-news-today-november-13-2026"&gt;AI news roundup for 13 November 2026&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why does this matter for India?
&lt;/h2&gt;

&lt;p&gt;Treat this section as analysis. Nvidia's filings do not break out India separately, so any Indian share of revenue is not disclosed. What is public is the dependence. Under the IndiaAI Mission, &lt;a href="https://pib.gov.in/PressReleaseIframePage.aspx?PRID=2245063&amp;amp;lang=2&amp;amp;reg=3" rel="noopener noreferrer"&gt;PIB reported&lt;/a&gt; 38,231 GPUs onboarded from 14 empanelled service providers, offered at roughly Rs 65 per GPU-hour to start-ups and academia - and PIB's own framing underlines the dependence: GPUs are &lt;a href="https://pib.gov.in/PressReleasePage.aspx?PRID=2245069&amp;amp;lang=2&amp;amp;reg=3" rel="noopener noreferrer"&gt;primarily manufactured in one country&lt;/a&gt;. Our guide to &lt;a href="https://dev.to/articles/agentic-ai-tutorial-indiaai-compute"&gt;agentic AI on IndiaAI compute&lt;/a&gt; covers how that capacity is being used.&lt;/p&gt;

&lt;p&gt;The permanent CEO will inherit the export-compliance posture that decides how quickly such capacity can grow. India's fab programme, covered in &lt;a href="https://dev.to/articles/semiconductor-industry-in-india-2026"&gt;semiconductor industry in India 2026&lt;/a&gt;, sits upstream at manufacturing rather than at the accelerator layer, so it does not remove that dependence. Developers choosing models to run on this hardware may also want our comparison of the &lt;a href="https://dev.to/articles/best-open-source-llm-for-coding-qwen3-coder-vs-glm-5-2-2026"&gt;best open-source LLM for coding&lt;/a&gt;. Investors weighing AI listings against this news can read &lt;a href="https://dev.to/articles/anthropic-ipo-vs-spacex-roadshow-70-78-ant-nasdaq"&gt;Anthropic IPO vs SpaceX roadshow&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: Who is the Nvidia interim CEO?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Colette Kress, previously executive vice president and CFO, was appointed interim CEO on 12 November 2026 and continues as CFO while the board searches for a permanent successor.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Why did Jensen Huang resign from Nvidia?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Nvidia's announcement does not give a reason. Reporting by AP and The New York Times links the resignation to allegations about Huang's personal conduct and a related national-security inquiry, which Nvidia said it was examining.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Will Nvidia's earnings date change?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. Nvidia said Q3 fiscal 2027 results remain scheduled for Tuesday 17 November, with the call at 2 p.m. PT.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Who is Nvidia's new chair?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Harvey C. Jones, a board member since 1993 and former CEO and Executive Chairman of Synopsys, becomes the company's first-ever Chair of the Board.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Will Colette Kress become the permanent CEO?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Not decided. The board said it will consider internal and external candidates, and analysts quoted by AP and The Guardian name both insiders and outsiders as possibilities.&lt;/p&gt;

</description>
      <category>nvidia</category>
      <category>jensenhuang</category>
      <category>colettekress</category>
      <category>semiconductors</category>
    </item>
    <item>
      <title>AI News Today: Lazarus Backdoor vs United Ban, Anthropic Cut</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Thu, 19 Nov 2026 04:43:57 +0000</pubDate>
      <link>https://dev.to/aitecharchive/ai-news-today-lazarus-backdoor-vs-united-ban-anthropic-cut-1f2h</link>
      <guid>https://dev.to/aitecharchive/ai-news-today-lazarus-backdoor-vs-united-ban-anthropic-cut-1f2h</guid>
      <description>&lt;p&gt;The biggest AI news today, 13 November 2026, is a three-way contest, and the winner for builders is Microsoft Research's Lazarus sleeper-agent backdoor. It needs fewer than 10 edits to a model's weights, no training, and it can sit dormant until a trigger appears. United Airlines' ban on third-party AI agents is the bigger consumer story, and Anthropic's new region rules matter if your company has any ownership link to mainland China, but only Lazarus changes what every team running agents should do this month.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Microsoft Research showed a backdoor that edits a handful of dormant neurons in an agent model. It is the story to act on: audit weights like code.&lt;/li&gt;
&lt;li&gt;United told third-party AI agents to stay out of its booking systems on 11 November, the clearest consumer-facing agent setback so far.&lt;/li&gt;
&lt;li&gt;Anthropic's Supported Regions Policy took effect on 12 November 2026 and now tests ownership, not just location.&lt;/li&gt;
&lt;li&gt;Google's Gemini Decisions API is the best new builder opportunity: &lt;a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/" rel="noopener noreferrer"&gt;$0.10 per million input tokens&lt;/a&gt;, no output tokens billed.&lt;/li&gt;
&lt;li&gt;Databricks repackaged Genie as an autonomous coworker for long-running enterprise tasks.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Last verified: 13 November 2026.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which story matters most in AI news today?
&lt;/h2&gt;

&lt;p&gt;Ranked by what you should do about it: Lazarus first, Anthropic's policy second, United third, with Gemini Decisions as the opportunity on the side.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Story&lt;/th&gt;
&lt;th&gt;Date&lt;/th&gt;
&lt;th&gt;Who it hits&lt;/th&gt;
&lt;th&gt;Your move&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Microsoft Lazarus backdoor&lt;/td&gt;
&lt;td&gt;Nov 10&lt;/td&gt;
&lt;td&gt;Any team deploying agents&lt;/td&gt;
&lt;td&gt;Audit model weights like code&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;United bans third-party AI agents&lt;/td&gt;
&lt;td&gt;Nov 11&lt;/td&gt;
&lt;td&gt;Consumer booking agents&lt;/td&gt;
&lt;td&gt;Build on official APIs, not logins&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Anthropic region rules&lt;/td&gt;
&lt;td&gt;Nov 12&lt;/td&gt;
&lt;td&gt;Firms with China-linked ownership&lt;/td&gt;
&lt;td&gt;Map the ownership chain now&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini Decisions API&lt;/td&gt;
&lt;td&gt;Nov 11&lt;/td&gt;
&lt;td&gt;Classification/routing workloads&lt;/td&gt;
&lt;td&gt;Prototype on the free tier&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Databricks Genie relaunch&lt;/td&gt;
&lt;td&gt;Nov 11&lt;/td&gt;
&lt;td&gt;Enterprise data teams&lt;/td&gt;
&lt;td&gt;Pilot with governed guardrails&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Lazarus is first because it is an attack on the model itself, not on the app around it. Anthropic's policy is second because it can switch off access for an organisation overnight, and the ownership test is easy to trip without realising. United is third for builders, though first for anyone building consumer booking agents, because it shows a large operator choosing to block agents rather than cooperate with them.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is the Microsoft Lazarus sleeper-agent backdoor?
&lt;/h2&gt;

&lt;p&gt;Lazarus is a technique, published by Microsoft Research on 10 November 2026, for planting a hidden behaviour in an AI agent by editing a few neurons that normally stay quiet. The &lt;a href="https://www.microsoft.com/en-us/research/blog/lazare-veni-foras-reprogramming-dormant-neurons-to-backdoor-ai-agents" rel="noopener noreferrer"&gt;Microsoft Research write-up&lt;/a&gt; and its &lt;a href="https://www.microsoft.com/en-us/security/blog/2026/11/10/detecting-sleeper-agents-hiding-in-ai-model-neurons" rel="noopener noreferrer"&gt;security blog companion&lt;/a&gt; describe the method, and the full paper is on &lt;a href="https://arxiv.org/pdf/2611.15118" rel="noopener noreferrer"&gt;arXiv&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The figures reported by Microsoft:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Fewer than 10 neuron edits, with no training and only inference-level compute.&lt;/li&gt;
&lt;li&gt;On Nemotron-Terminal-8B, 7 of more than 12,000 final-layer neurons were changed (&lt;a href="https://arxiv.org/pdf/2611.15118" rel="noopener noreferrer"&gt;arXiv paper&lt;/a&gt;), fitted using 23 trajectories.&lt;/li&gt;
&lt;li&gt;The payload fired on at least 98.8 percent of triggered test trajectories, with no observed leakage on benign ones.&lt;/li&gt;
&lt;li&gt;The approach carried over to Qwen 3.6-27B and Gemma-4-31B-IT.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The trigger can be something as ordinary as a date rollover (the example given is 2027-01-01) or a particular string in a tool's output. The attacker only needs write access to the weights, so a model you trained yourself is as exposed as a downloaded one. Microsoft states that damage scales with the privileges the agent holds, not with the size of the edit, and that it has seen no threat actors using the technique.&lt;/p&gt;

&lt;p&gt;On defence, Microsoft reports that agent-built detectors with no reference model flagged 91.4 percent of edited checkpoints, at a 9.7 percent false-positive rate. That is useful, not sufficient. One in ten clean checkpoints flagged is a lot of noise for a release pipeline.&lt;/p&gt;

&lt;h2&gt;
  
  
  What should teams do about Lazarus this week?
&lt;/h2&gt;

&lt;p&gt;Treat weights as a supply-chain artefact. In practice:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Record the source and hash of every checkpoint you deploy, and verify it at load time.&lt;/li&gt;
&lt;li&gt;Restrict who can write to model storage, the same way you restrict production deploy keys.&lt;/li&gt;
&lt;li&gt;Reduce agent privileges. A backdoored agent with read-only access can do far less than one that can run shell commands or spend money.&lt;/li&gt;
&lt;li&gt;Test agents against date changes and unusual tool output before the year ends, since the published example trigger is a new-year rollover.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;For related reading on agent risk, our piece on &lt;a href="https://dev.to/articles/gpt-6-cyber-vs-gpt-6-astra-exploit-verdict-2026"&gt;GPT-6 Cyber vs GPT-6 Astra&lt;/a&gt; covers the offensive side of security tooling, and the &lt;a href="https://dev.to/articles/deepseek-harness-dsh-open-source-coding-agent-leaderboard"&gt;DeepSeek Harness leaderboard&lt;/a&gt; is worth a look if you run open-weight coding agents, which are the models most often fetched from third-party sources.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why did United ban third-party AI agents?
&lt;/h2&gt;

&lt;p&gt;United says unauthorised personal AI agents create risks around credential exposure, unintended purchases, account changes and misuse of MileagePlus. In its &lt;a href="https://www.united.com/en/us/newsroom/announcements/cision-125496" rel="noopener noreferrer"&gt;11 November statement&lt;/a&gt;, the airline said it is restricting access both directly on its platforms and indirectly through travel agencies and aggregators, and is updating booking, ticketing and distribution policies.&lt;/p&gt;

&lt;p&gt;It also pointed to regulatory duties: flight check-in requires a human to acknowledge that bags contain no banned substances such as explosives, which an agent cannot truthfully do on someone's behalf. The statement notes that banking and retail leaders have taken similar steps.&lt;/p&gt;

&lt;p&gt;For anyone building a travel or shopping agent, the practical reading is that scraping or automating a logged-in session is now a contractual and technical risk. The sturdier route is an official API or a partner agreement, where the operator keeps control of the human-acknowledgement steps.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does Anthropic's new region policy affect your team?
&lt;/h2&gt;

&lt;p&gt;Anthropic's &lt;a href="https://www.anthropic.com/legal/supported-regions" rel="noopener noreferrer"&gt;Supported Regions Policy&lt;/a&gt;, effective 12 November 2026, excludes anyone who meets any of three tests: they are physically in an unsupported region, their entity is incorporated or headquartered in one, or their entity is majority-owned or controlled, directly or indirectly, by people or entities in one. The third test applies wherever the staff sit.&lt;/p&gt;

&lt;p&gt;The supported list omits mainland China, Hong Kong and Macau. Taiwan and India are listed. The policy covers Claude.ai, Claude Code, Cowork, the API and developer platform, and cloud platforms and authorised resellers.&lt;/p&gt;

&lt;p&gt;The action for compliance teams is to map your ownership chain, not your office locations. A London or Bangalore subsidiary of a group controlled from an unsupported region can fall outside the policy. If you depend on Claude Code in your workflow, see our notes on &lt;a href="https://dev.to/articles/claude-code-2-1-323-code-mode-vs-standard-tool-calls"&gt;Claude Code 2.1.323&lt;/a&gt; and &lt;a href="https://dev.to/articles/claude-code-2-1-321-fable-5-2-default-price-security"&gt;Claude Code 2.1.321&lt;/a&gt; for the tooling side, and plan a fallback provider in case access changes.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is Gemini Decisions API worth trying?
&lt;/h2&gt;

&lt;p&gt;For classification-heavy workloads, yes. Google DeepMind's &lt;a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/" rel="noopener noreferrer"&gt;announcement&lt;/a&gt; describes a model that scores up to 1,000 yes/no, multiple-choice or rubric questions in one forward pass over a single shared multimodal input of up to 256K tokens. It returns probabilities and confidence scores rather than generated text.&lt;/p&gt;

&lt;p&gt;Google reports 150 to 250 ms end-to-end latency, nearly flat from 1 to 100 questions, and a price of &lt;a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/" rel="noopener noreferrer"&gt;$0.10 per million input tokens&lt;/a&gt; with no output tokens billed. It also says the service is &lt;a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/" rel="noopener noreferrer"&gt;nearly 5x cheaper&lt;/a&gt; than Gemini 3.5 Flash-Lite with the Interactions API on a shared benchmark suite. That is Google's own comparison, so test it on your data. There is a free tier in Google AI Studio, and launch integrations include OpenRouter, LangChain, Vercel and LiteLLM.&lt;/p&gt;

&lt;p&gt;The limitation is obvious: it judges, it does not write. It suits moderation, routing, grading and triage, not drafting.&lt;/p&gt;

&lt;h2&gt;
  
  
  What did Databricks announce with Genie?
&lt;/h2&gt;

&lt;p&gt;Databricks &lt;a href="https://www.databricks.com/blog/new-genie-autonomous-ai-complex-enterprise-work" rel="noopener noreferrer"&gt;relaunched Genie&lt;/a&gt; as an autonomous coworker for long-running enterprise tasks, with Genie Ontology and Genie Search generally available and admin controls in Unity Gateway covering tool access, action tracing and spend limits. Databricks says that in a benchmark built with Turing, Genie with Ontology passed 91 percent of enterprise task runs, against 45 and 58 percent for two frontier agents, using less than half the tokens. This is a vendor benchmark, so treat it as a claim to reproduce, not a ranking. For the labour-market context, see &lt;a href="https://dev.to/articles/gartner-ai-jobs-forecast-vs-qantas-ai-washing-2026"&gt;Gartner's AI jobs forecast vs Qantas&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: What is the most important AI news today?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Microsoft Research's Lazarus backdoor, published on 10 November 2026, because it shows an agent model can be compromised with fewer than 10 neuron edits and no training, and it affects any team deploying agents.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Has United banned all AI agents?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It has banned unauthorised third-party personal AI agents and assistants from its digital platforms and, indirectly, from travel agencies and aggregators, according to its 11 November 2026 statement.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does Anthropic's policy block users in India?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. India is on Anthropic's supported list. The policy bars mainland China, Hong Kong and Macau, and entities owned or controlled from unsupported regions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is the Lazarus backdoor being used in real attacks?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Microsoft says it has observed no threat actors weaponising it. It was published as research, along with a detection approach.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How much does the Gemini Decisions API cost?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Google lists &lt;a href="https://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/" rel="noopener noreferrer"&gt;$0.10 per million input tokens&lt;/a&gt; with no output tokens billed, plus a no-cost tier in Google AI Studio.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;This digest was compiled with AI assistance and checked against the primary sources linked above. If you spot an error, tell us and we will log a correction.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>ainewstoday</category>
      <category>aiagents</category>
      <category>aisecurity</category>
      <category>anthropic</category>
    </item>
    <item>
      <title>Cursor Projects Multiplayer vs Antigravity Swarm</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Thu, 19 Nov 2026 04:43:37 +0000</pubDate>
      <link>https://dev.to/aitecharchive/cursor-projects-multiplayer-vs-antigravity-swarm-5bcl</link>
      <guid>https://dev.to/aitecharchive/cursor-projects-multiplayer-vs-antigravity-swarm-5bcl</guid>
      <description>&lt;p&gt;Cursor Projects multiplayer, announced on 10 November 2026 in &lt;a href="https://cursor.com/blog/projects-multiplayer" rel="noopener noreferrer"&gt;Cursor's own post&lt;/a&gt;, lets the coding agents of different teammates coordinate inside one shared Project. Google's Antigravity Swarm, by contrast, organises the agents working for a single developer. Our verdict: for a team on a paid Cursor plan, Cursor Projects multiplayer is the more useful of the two because it solves handoffs and overlapping work between people. For a solo developer, Swarm is the only one of the two that does anything for you, and Cursor's release adds nothing.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Cursor Projects multiplayer is in beta and began rolling out to all Cursor users on 10 November 2026, per the &lt;a href="https://cursor.com/blog/projects-multiplayer" rel="noopener noreferrer"&gt;announcement&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Each teammate who joins gets their own chat, coordinator and worker agents, billed to that person.&lt;/li&gt;
&lt;li&gt;Members share two things only: the task board and one shared Context store.&lt;/li&gt;
&lt;li&gt;Messages between teammates' agents are treated as untrusted input, and the side chats that answer them are deliberately restricted.&lt;/li&gt;
&lt;li&gt;Antigravity Swarm works within one developer's assignment; Cursor works across teammates' assignments.&lt;/li&gt;
&lt;li&gt;Pilot it on one bounded workflow, such as an on-call rotation. Free and Plus users cannot use it.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What changed with Cursor Projects multiplayer?
&lt;/h2&gt;

&lt;p&gt;Projects first launched in September 2026 as a single-engineer feature: one person directs a fleet of agents through a coordinator that plans and delegates but does not write code itself. Projects run on Cloud Agents, so closing your laptop does not stop the work, per the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Projects documentation&lt;/a&gt;. Our piece on &lt;a href="https://dev.to/articles/cursor-live-apps-cloud-agents-hosted-web-apps"&gt;Cursor Live Apps and cloud agents&lt;/a&gt; covers the hosting side.&lt;/p&gt;

&lt;p&gt;The November change is that you can invite teammates into the same Project. Cursor's product manager Aryaman Khandelwal framed it in the announcement: multi-agent systems have so far served one person; Projects is now built for a whole team. The forum thread &lt;a href="https://forum.cursor.com/t/multiplayer-cursor-projects/175478" rel="noopener noreferrer"&gt;"Multiplayer Cursor Projects"&lt;/a&gt; shows users had been asking for this, since teammates' agents could not see each other. Independent write-ups from &lt;a href="https://unite.ai/cursor-adds-team-collaboration-and-agent-handoffs-to-projects" rel="noopener noreferrer"&gt;unite.ai&lt;/a&gt; and &lt;a href="https://ai.info/news/cursor-projects-shared-team-workspaces-beta" rel="noopener noreferrer"&gt;ai.info&lt;/a&gt; match the announcement on the rollout date, the shared task board and the team-account gating.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does a shared Project actually work?
&lt;/h2&gt;

&lt;p&gt;Each person who joins gets their own chat inside the Project, with their own coordinator and worker agents. That chat runs and bills as that person, not as the Project owner, according to the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;docs&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;What is shared is deliberately narrow:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The task board.&lt;/strong&gt; Everyone sees the same tasks in the Project's list and board.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Context.&lt;/strong&gt; One Agent Store, mounted at &lt;code&gt;/cursor/stores/self&lt;/code&gt;, that every member's agents can read and write.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Your personal store at &lt;code&gt;/cursor/stores/user&lt;/code&gt; stays private. Only your own agents mount it, and it holds your preferences and skills. When a teammate joins, your coordinator can offer to copy approved entries into Project rules, so shared conventions are an explicit choice rather than a leak.&lt;/p&gt;

&lt;h2&gt;
  
  
  What do you need before you can invite teammates?
&lt;/h2&gt;

&lt;p&gt;The docs list real gating conditions - check these before planning a pilot.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Requirement&lt;/th&gt;
&lt;th&gt;Detail&lt;/th&gt;
&lt;th&gt;Source&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Account type&lt;/td&gt;
&lt;td&gt;A team account&lt;/td&gt;
&lt;td&gt;&lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Docs&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cursor version&lt;/td&gt;
&lt;td&gt;3.26.13 or later for sharing a Project; Enterprise teams need 3.21.9 or later for base Projects&lt;/td&gt;
&lt;td&gt;&lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Docs&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Plan&lt;/td&gt;
&lt;td&gt;Any plan that includes Cloud Agents: Premium and higher, Teams, Enterprise, legacy paid plans&lt;/td&gt;
&lt;td&gt;&lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Docs&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Repository access&lt;/td&gt;
&lt;td&gt;Each invitee must have access to the Project's repository&lt;/td&gt;
&lt;td&gt;&lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Docs&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Privacy&lt;/td&gt;
&lt;td&gt;Joining is unavailable with Privacy Mode (Legacy)&lt;/td&gt;
&lt;td&gt;&lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Docs&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Project state&lt;/td&gt;
&lt;td&gt;Not available on a self-hosted machine or an archived Project&lt;/td&gt;
&lt;td&gt;&lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;Docs&lt;/a&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Free, Plus and legacy Hobby plans are not covered. If you are weighing plans, the comparison in &lt;a href="https://dev.to/articles/google-antigravity-vs-cursor-vs-claude-code-2026"&gt;Google Antigravity vs Cursor vs Claude Code&lt;/a&gt; is a sensible companion read.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do teammates' agents coordinate in practice?
&lt;/h2&gt;

&lt;p&gt;Cursor gives two examples: an engineer adding reviews to a product page while a designer redesigns it, whose coordinators spot the overlap and sync so the reviews land in the new design; and a feature reading a database while a teammate changes the schema, where agents agree on work order up front to avoid merge conflicts.&lt;/p&gt;

&lt;p&gt;The second is more convincing: schema-versus-reader conflicts are tedious to spot in review and costly to untangle after both branches exist.&lt;/p&gt;

&lt;p&gt;The on-call case is specific: several Cursor teams run rotations in a shared Project, and an engineer taking over finds the coordinator already knows work in flight, unattended issues, incidents and notes needing handoff, plus channels, Linear boards and threads to watch. In-progress work can be handed over directly.&lt;/p&gt;

&lt;p&gt;We have not run this ourselves; these are vendor-described scenarios and the feature is in beta.&lt;/p&gt;

&lt;h2&gt;
  
  
  Can agents from different teammates talk to each other safely?
&lt;/h2&gt;

&lt;p&gt;Cursor's answer is to restrict what the receiving side can do. Per the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;docs&lt;/a&gt;, a message from a teammate's coordinator is labelled as agent input. The receiving agent is told which teammate's agent sent it, treats the text as untrusted, and asks before sharing that person's personal files or connector data.&lt;/p&gt;

&lt;p&gt;The message lands in a side chat forked from your chat, named "Chat with {name}'s agent". It answers from what your chat already knows, the reply goes back to the sender, and your chat gets a note on its next turn.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Side chat can&lt;/th&gt;
&lt;th&gt;Side chat cannot&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Answer from your chat's context&lt;/td&gt;
&lt;td&gt;Share secrets, credentials or tokens&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Send the reply back to the sender's chat&lt;/td&gt;
&lt;td&gt;Change the task board&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Read the task board&lt;/td&gt;
&lt;td&gt;Ask questions or request access&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Be read, stopped or archived by you&lt;/td&gt;
&lt;td&gt;Connect source control or manage routines&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;You cannot message a side chat; you can only read, stop or archive it. This is a sound default: agent-to-agent text is an injection surface, and documenting what the responder cannot do beats a general safety promise. It does not remove the need to review what lands in shared Context, because every member's agents can write there.&lt;/p&gt;

&lt;h2&gt;
  
  
  Cursor Projects multiplayer or Antigravity Swarm: which should you use?
&lt;/h2&gt;

&lt;p&gt;The two tools sit at different layers. According to &lt;a href="https://thenewstack.io/cursor-google-multiplayer-swarm" rel="noopener noreferrer"&gt;The New Stack's coverage&lt;/a&gt;, Swarm works within one developer's assignment, while multiplayer Projects works across teammates' assignments. &lt;a href="https://antigravity.google/docs/swarm/" rel="noopener noreferrer"&gt;Google's Swarm documentation&lt;/a&gt; describes the single-developer model. The same report puts Swarm in preview for Google AI Pro and Ultra subscribers, and multiplayer Projects in beta for team-account users.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Cursor Projects multiplayer&lt;/th&gt;
&lt;th&gt;Antigravity Swarm&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Scope&lt;/td&gt;
&lt;td&gt;Across teammates&lt;/td&gt;
&lt;td&gt;One developer&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Status&lt;/td&gt;
&lt;td&gt;Beta, team accounts&lt;/td&gt;
&lt;td&gt;Preview, Google AI Pro and Ultra&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best for&lt;/td&gt;
&lt;td&gt;Handoffs, overlapping work&lt;/td&gt;
&lt;td&gt;Parallelising your own task&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Weak spot&lt;/td&gt;
&lt;td&gt;Needs plan, version and repo access&lt;/td&gt;
&lt;td&gt;No cross-person coordination&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The verdict follows from that split. A team sharing a codebase gets more from Cursor, because the expensive failures are between people. A solo developer gets nothing from Cursor's release and should look at Swarm or at the options in our &lt;a href="https://dev.to/articles/claude-code-2-1-323-code-mode-vs-standard-tool-calls"&gt;Claude Code 2.1.323 analysis&lt;/a&gt;. For a wider view of Cursor's direction, see our &lt;a href="https://dev.to/articles/cursor-origin-review"&gt;Cursor Origin review&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;Context: Cursor &lt;a href="https://cursor.com/blog/joining-spacex" rel="noopener noreferrer"&gt;now belongs to SpaceX&lt;/a&gt;, as the company announced in August 2026 — that affects model supply more than this feature, so we would not weigh it heavily in a tooling decision. The wider trend matters more: multi-agent coding is moving from personal tooling to team infrastructure, and agent identity, attribution and untrusted-input handling are now product features rather than afterthoughts.&lt;/p&gt;

&lt;p&gt;Within our 72 winnable keywords, 20 carry a difficulty score of zero, meaning no established competitor holds the result set (measured 2026-11-12, n=72). This feature is one of those gaps, so most results today are the vendor's own pages. Treat third-party summaries with the same caution and read the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;docs&lt;/a&gt; directly.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is a sensible way to pilot it?
&lt;/h2&gt;

&lt;p&gt;Pick one bounded workflow and one small group.&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Choose an on-call rotation or migration handoff, where continuity matters and scope is clear.&lt;/li&gt;
&lt;li&gt;Confirm each invitee is on a team account, on Cursor 3.26.13 or later, and has repository access.&lt;/li&gt;
&lt;li&gt;Decide which conventions go into Project rules and which stay in personal stores.&lt;/li&gt;
&lt;li&gt;After the first week, review what agents wrote to shared Context and compare handoff time and merge conflicts against your previous rotation.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Skip it if you work alone or your plan lacks Cloud Agents. VS Code users may find our note on &lt;a href="https://dev.to/articles/vscode-1-146-inbox-mode-fusion-discount-deadline"&gt;VS Code 1.146 Inbox mode&lt;/a&gt; relevant.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: What is Cursor Projects multiplayer?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A:&lt;/strong&gt; A beta feature, rolling out from 10 November 2026, that lets teammates join one Cursor Project; each gets own coordinator and worker agents, while the task board and a shared Context store are common to all, per the &lt;a href="https://cursor.com/blog/projects-multiplayer" rel="noopener noreferrer"&gt;announcement&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Who can join a shared Project?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A:&lt;/strong&gt; Teammates on a team account running Cursor 3.26.13 or later, with access to the Project's repository, as set out in the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;docs&lt;/a&gt;. Joining is unavailable with Privacy Mode (Legacy), on a self-hosted Project, or on an archived Project.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is Cursor Projects multiplayer free?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A:&lt;/strong&gt; No. Projects need a plan that includes Cloud Agents — Premium and higher, Teams or Enterprise — and are unavailable on Free, Plus or legacy Hobby, per the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;docs&lt;/a&gt;. Each member's chat is billed to that member.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How do agents from different teammates communicate safely?&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A:&lt;/strong&gt; Incoming messages are labelled with the sender and treated as untrusted input. They are answered in a restricted side chat that cannot share secrets, change the task board, ask questions or manage routines, per the &lt;a href="https://cursor.com/docs/agent/projects" rel="noopener noreferrer"&gt;docs&lt;/a&gt;.&lt;/p&gt;

</description>
      <category>cursor</category>
      <category>multiagent</category>
      <category>antigravity</category>
      <category>cloudagents</category>
    </item>
    <item>
      <title>DeepSeek Harness vs Hermes Agent: Tied First, One Flaw</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Thu, 19 Nov 2026 04:43:17 +0000</pubDate>
      <link>https://dev.to/aitecharchive/deepseek-harness-vs-hermes-agent-tied-first-one-flaw-3klp</link>
      <guid>https://dev.to/aitecharchive/deepseek-harness-vs-hermes-agent-tied-first-one-flaw-3klp</guid>
      <description>&lt;p&gt;DeepSeek Harness (DSH) is the better pick for repository code changes on open models, and Hermes Agent, Pi or OpenCode is the better pick when background services must keep running after the agent exits. The &lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard" rel="noopener noreferrer"&gt;Harness Leaderboard&lt;/a&gt; added DSH on 9 November 2026 and, per its &lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard/blob/main/CHANGELOG.md" rel="noopener noreferrer"&gt;changelog&lt;/a&gt;, ranked it in every view on 11 November. It is statistically tied for first in all five evaluations, yet it places only 6th of 10 on Terminal-Bench 2.1 because of a single cleanup habit.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;DSH is tied for first in all five board evaluations, where "tied" means the 95% bootstrap interval of the paired difference to the leader covers zero (&lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;Niklaus announcement&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;Averaged over three models, DSH is 2nd on DeepSWE v1.1, 0.5 points behind Hermes Agent, and 6th on Terminal-Bench 2.1.&lt;/li&gt;
&lt;li&gt;The gap comes from DSH shutting down every background program it started when a task ends. Three Terminal-Bench tasks grade a server after the agent exits.&lt;/li&gt;
&lt;li&gt;DSH averaged $0.53 per solved trial on the board.&lt;/li&gt;
&lt;li&gt;The harness is MIT-licensed and still a developer preview, so expect breaking changes.&lt;/li&gt;
&lt;li&gt;Last verified: 13 November 2026.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How does DeepSeek Harness compare on the Harness Leaderboard?
&lt;/h2&gt;

&lt;p&gt;The board, run by Hugging Face researcher Joël Niklaus, holds the model and reasoning effort fixed and swaps only the harness. It scores Terminal-Bench 2.1 and DeepSWE v1.1 with three attempts per task, according to the &lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard" rel="noopener noreferrer"&gt;Space&lt;/a&gt;. The five evaluations are DeepSeek-V4.1-Flash and GLM-5.3-Flash at low and high effort, plus Flint.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Where&lt;/th&gt;
&lt;th&gt;Result&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Five board evaluations&lt;/td&gt;
&lt;td&gt;Statistically tied for first in all 5&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;DeepSWE, averaged over 3 models&lt;/td&gt;
&lt;td&gt;2nd of 10, 0.5 points behind Hermes Agent&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Terminal-Bench 2.1, averaged&lt;/td&gt;
&lt;td&gt;6th of 10&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GLM-5.3-Flash at high effort&lt;/td&gt;
&lt;td&gt;5th of 10, tied with leader&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Source for the table: the &lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard/blob/main/CHANGELOG.md" rel="noopener noreferrer"&gt;board changelog&lt;/a&gt; and &lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;Niklaus's announcement&lt;/a&gt;. A 5th-place rank that is still tied with the leader is a reminder that the field is tight. Before DSH arrived, the top five harnesses sat within 2.2 points of each other, according to &lt;a href="https://codemago.dev/news/en/news/harness-leaderboard-swapping-harness-moves-ai-score-24-points" rel="noopener noreferrer"&gt;CodeMago's write-up&lt;/a&gt; of the launch week.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why does DeepSeek Harness lose ground on Terminal-Bench?
&lt;/h2&gt;

&lt;p&gt;DSH shuts down every background program it started once it finishes a task, as &lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;Niklaus reports&lt;/a&gt;. Three Terminal-Bench tasks ask the agent to start a server, for example a model inference API, and the grader then tests that server after the agent has exited. DSH passed 2 of 36 attempts on those tasks, while OpenCode, Pi and Hermes Agent passed 36 of 36.&lt;/p&gt;

&lt;p&gt;Those three tasks cost DSH between 1.5 and 1.7 points per evaluation. That is more than its 1.3-point gap to the leader on DeepSeek-V4.1-Flash at high effort (&lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;announcement&lt;/a&gt;). In other words, the flaw is a behavioural default, not a lack of capability, and it is the whole reason DSH does not top the Terminal-Bench view.&lt;/p&gt;

&lt;h2&gt;
  
  
  What can you do if your server must outlive the agent?
&lt;/h2&gt;

&lt;p&gt;This section is our own analysis, not a claim from the board or its author. Given the confirmed behaviour, there are two sensible workarounds:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Start long-lived services under &lt;code&gt;systemd&lt;/code&gt; or with &lt;code&gt;nohup&lt;/code&gt; and &lt;code&gt;setsid&lt;/code&gt;, so they sit outside the agent's own process group and are not swept up at cleanup.&lt;/li&gt;
&lt;li&gt;State the persistence requirement explicitly in the prompt, for instance "leave the server running after you finish".&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;We have not tested either against DSH, so treat them as a starting point and verify with a quick task before relying on them. If a workflow depends on this often, the simpler route is to choose a harness that already passes this class of task.&lt;/p&gt;

&lt;h2&gt;
  
  
  Is DeepSeek Harness cheap to run?
&lt;/h2&gt;

&lt;p&gt;DSH averaged $0.53 per solved trial on the board (&lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;Niklaus&lt;/a&gt;). The board lists DeepSeek-V4.1-Flash at $0.28 per million input tokens and $1.12 per million output tokens, and GLM-5.3-Flash at $0.15 per million input tokens (&lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard" rel="noopener noreferrer"&gt;Space&lt;/a&gt;). The harness itself costs nothing, so your bill is the model API.&lt;/p&gt;

&lt;p&gt;One caveat on scoring: 20 of DSH's passes were overturned, for example for fetching upstream code, which the board counts as a fail (&lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard" rel="noopener noreferrer"&gt;board method&lt;/a&gt;). DSH also ran with web search off, because its search needs a DeepSeek account key the test machines do not hold.&lt;/p&gt;

&lt;h2&gt;
  
  
  Does the harness matter as much as the model?
&lt;/h2&gt;

&lt;p&gt;Independent research suggests it does. An &lt;a href="https://arxiv.org/abs/2610.00917" rel="noopener noreferrer"&gt;arXiv study of 66 model-harness configurations&lt;/a&gt; found model rankings reverse across harnesses. On Terminal-Bench 4, Claude leads GPT by 7.94 points in OpenHands but trails it by 30.16 points in PI. The same paper reports DSH as the best harness for GLM-5.3 and DeepSeek V4 Pro on that benchmark.&lt;/p&gt;

&lt;p&gt;The board's own launch data points the same way: swapping the harness moved scores by up to 24 points for weaker harnesses, per &lt;a href="https://codemago.dev/news/en/news/harness-leaderboard-swapping-harness-moves-ai-score-24-points" rel="noopener noreferrer"&gt;CodeMago&lt;/a&gt;. For more on how agent platforms differ in practice, see our &lt;a href="https://dev.to/articles/best-open-source-ai-agent-platform-for-coding-in-2026-openmausbot-vs-alternatives"&gt;open-source coding agent platform comparison&lt;/a&gt; and the &lt;a href="https://dev.to/articles/claude-code-2-1-323-code-mode-vs-standard-tool-calls"&gt;Claude Code tool-calling comparison&lt;/a&gt;. For the model side of the same question, read &lt;a href="https://dev.to/articles/swe-2-flash-vs-glm-5-3-flash-compared"&gt;SWE-2 Flash vs GLM-5.3 Flash&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you install and run DeepSeek Harness?
&lt;/h2&gt;

&lt;p&gt;These steps come from the &lt;a href="https://github.com/deepseek-ai/deepseek-harness" rel="noopener noreferrer"&gt;DeepSeek Harness repository&lt;/a&gt;.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Install Node.js, version 22.19 or later, or 24 and above.&lt;/li&gt;
&lt;li&gt;Run &lt;code&gt;npx @deepseek-ai/dsh web&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Open the Web UI at &lt;code&gt;http://127.0.0.1:3080&lt;/code&gt; and add a DeepSeek API key under Settings, then Models.&lt;/li&gt;
&lt;li&gt;Start a session in the folder you want DSH to work in.&lt;/li&gt;
&lt;li&gt;To build from source, you need Node ^22.19 or 24+ and pnpm 11, then run &lt;code&gt;pnpm install &amp;amp;&amp;amp; pnpm run build &amp;amp;&amp;amp; pnpm dsh web&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Session logs stay local by default. Setting &lt;code&gt;DSH_TELEMETRY_MODE=FEEDBACK_ONLY&lt;/code&gt; shares a log only when you submit feedback, and &lt;code&gt;FULL&lt;/code&gt; uploads continuously. The repository carries a SAFETY.md notice that you should read before running an agent with file and shell access. DSH is a developer preview (the board tested 0.2.2-rc.1), and the repo warns that compatibility-breaking changes are expected. Its architecture treats everything as a plugin on Cordis, and profiles are ordered lists of plugin bundles managed with &lt;code&gt;dsh plugin&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The project is MIT-licensed, with 279,576 stars and 34,424 forks at the time of checking, created on 13 August 2026 (&lt;a href="https://github.com/deepseek-ai/deepseek-harness" rel="noopener noreferrer"&gt;GitHub&lt;/a&gt;). Popularity is not evidence of quality, which is why the independent board result matters more.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which harness should you choose?
&lt;/h2&gt;

&lt;p&gt;Choose DSH when the job is repository code changes on open models, the kind of work DeepSWE measures. Choose Hermes Agent, Pi or OpenCode when servers, watchers or dev APIs must survive after the agent finishes. Whichever you pick, verify results before trusting them, as we argue in &lt;a href="https://dev.to/articles/agentic-engineering-verify-first-vs-build-first"&gt;verify-first vs build-first agentic engineering&lt;/a&gt;, and plan the task before handing it over, as in &lt;a href="https://dev.to/articles/shadcn-improve-10423-stars-plan-first-coding"&gt;this plan-first coding workflow&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The board is also still moving. Cline is listed as the next harness in preparation (&lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard/blob/main/CHANGELOG.md" rel="noopener noreferrer"&gt;changelog&lt;/a&gt;), so rankings may shift. For a week-one view of the board's totals and caveats, &lt;a href="https://bokonon.ai/notes/2026-11-06-a-tie-a-tail-and-a-cap/" rel="noopener noreferrer"&gt;bokonon.ai's analysis&lt;/a&gt; is worth reading.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: What is DeepSeek Harness?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; DeepSeek Harness (DSH) is an open-source coding agent from DeepSeek AI, released under the MIT licence. It uses an "everything is a plugin" architecture built on Cordis, according to its &lt;a href="https://github.com/deepseek-ai/deepseek-harness" rel="noopener noreferrer"&gt;GitHub repository&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is DeepSeek Harness free?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The harness is free and open source, but you pay for the model API. The Harness Leaderboard measured an average of $0.53 per solved trial (&lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;Niklaus&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Why does DeepSeek Harness fail some Terminal-Bench tasks?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It shuts down background programs it started when it finishes, while three tasks grade a server after the agent exits. DSH passed 2 of 36 attempts on them, per the &lt;a href="https://x.com/joelniklaus/status/2120540528886603949" rel="noopener noreferrer"&gt;announcement&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Which models can run in DeepSeek Harness?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It talks to models through an OpenAI-compatible endpoint. The board used DeepSeek-V4.1-Flash, GLM-5.3-Flash and Flint, per the &lt;a href="https://huggingface.co/spaces/joelniklaus/harness-leaderboard" rel="noopener noreferrer"&gt;Space&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Should I use DeepSeek Harness or Hermes Agent, Pi or OpenCode?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Use DSH for repository code changes. Use Hermes Agent, Pi or OpenCode when background services must stay alive after the agent exits.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does DeepSeek Harness send my data anywhere?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Not by default. Session logs stay local unless you set &lt;code&gt;DSH_TELEMETRY_MODE&lt;/code&gt; to &lt;code&gt;FEEDBACK_ONLY&lt;/code&gt; or &lt;code&gt;FULL&lt;/code&gt;, as described in the &lt;a href="https://github.com/deepseek-ai/deepseek-harness" rel="noopener noreferrer"&gt;repository&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Corrections: none yet. This article was researched and drafted with AI assistance and edited by a human; see our editorial policy for details.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>deepseekharness</category>
      <category>codingagents</category>
      <category>harnessleaderboard</category>
      <category>hermesagent</category>
    </item>
    <item>
      <title>Claude Team Plan Price vs Pro: Team Wins for Meetings</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Wed, 18 Nov 2026 04:44:55 +0000</pubDate>
      <link>https://dev.to/aitecharchive/claude-team-plan-price-vs-pro-team-wins-for-meetings-21mk</link>
      <guid>https://dev.to/aitecharchive/claude-team-plan-price-vs-pro-team-wins-for-meetings-21mk</guid>
      <description>&lt;p&gt;The Claude Team plan price in India is Rs 2,399 per seat per month on annual billing (&lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;), so a three-seat workspace costs Rs 7,197 a month. Three separate Claude Pro plans cost Rs 6,000 a month on annual billing (&lt;a href="https://moneycontrol.com/technology/claude-india-pricing-for-pro-max-and-team-subscriptions-revealed-here-s-how-much-anthropic-s-ai-plans-cost-article-13972503.html" rel="noopener noreferrer"&gt;Moneycontrol&lt;/a&gt;), which makes Team Rs 1,197 a month dearer. Team wins if your small business runs on meetings, because Claude Meetings is limited to Team and Enterprise plans; three Pro seats remain the better buy if you work alone or only use chat.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Claude Team Standard costs Rs 2,399 per seat per month on annual billing in India, according to &lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Three Team seats come to Rs 7,197 a month; three Pro seats on annual billing come to Rs 6,000 (&lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;Claude Whiteboards and Claude Sheets (both beta) and Claude Data (now generally available) arrived on 10 November 2026 for all paid plans.&lt;/li&gt;
&lt;li&gt;Claude Meetings, in beta from 12 November 2026, is for Team and Enterprise only, and works with Google Meet first.&lt;/li&gt;
&lt;li&gt;Choose Team if meetings and shared boards are daily work. Choose Pro x3 if each person works alone in chat.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What did Anthropic add to Claude in November 2026?
&lt;/h2&gt;

&lt;p&gt;Anthropic shipped two launches within roughly 72 hours. On 10 November it released Claude Whiteboards and Claude Sheets in beta on all paid plans, and moved Claude Data (previously Claude Dashboards) to general availability, as described in its &lt;a href="https://claude.com/resources/articles/sheets-whiteboards-data" rel="noopener noreferrer"&gt;announcement&lt;/a&gt;. &lt;a href="https://techcrunch.com/2026/11/10/anthropic-launches-claude-meetings-to-turn-workplace-discussions-into-plans-and-tasks" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt; and &lt;a href="https://www.engadget.com/2312177/anthropic-takes-on-google-and-microsoft-with-claude-sheets" rel="noopener noreferrer"&gt;Engadget&lt;/a&gt; reported the same event.&lt;/p&gt;

&lt;p&gt;On 12 November it followed with &lt;a href="https://claude.com/resources/articles/claude-meetings" rel="noopener noreferrer"&gt;Claude Meetings&lt;/a&gt;, where Claude joins a call as a visible participant and builds documents, whiteboards and notes while people talk. OpenAI shipped its own office-style tools in ChatGPT on 29 September 2026, so this is a direct response to a competitor, not an isolated feature drop.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is the Claude Team plan price in India, and how does it compare with Pro?
&lt;/h2&gt;

&lt;p&gt;Anthropic introduced rupee pricing for Pro, Max and Team on 13 July 2026, per &lt;a href="https://techcrunch.com/2026/07/13/anthropic-starts-localizing-claude-pricing-for-india-its-biggest-market-after-the-us" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;. Team Standard is Rs 2,399 per seat per month on annual billing, per &lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;. In US dollars &lt;a href="https://techcrunch.com/2026/11/10/anthropic-launches-claude-meetings-to-turn-workplace-discussions-into-plans-and-tasks" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt; puts Team at a $25 to $30 per seat range. Pro in India is Rs 2,000 a month on annual billing and Rs 2,399 on monthly billing, as reported by &lt;a href="https://moneycontrol.com/technology/claude-india-pricing-for-pro-max-and-team-subscriptions-revealed-here-s-how-much-anthropic-s-ai-plans-cost-article-13972503.html" rel="noopener noreferrer"&gt;Moneycontrol&lt;/a&gt; and covered in our &lt;a href="https://dev.to/articles/claude-ai-pro-price-in-india-vs-google-ai-pro"&gt;Claude Pro price comparison with Google AI Pro&lt;/a&gt;.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Three people&lt;/th&gt;
&lt;th&gt;Monthly cost&lt;/th&gt;
&lt;th&gt;What you get&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Claude Team, annual billing&lt;/td&gt;
&lt;td&gt;Rs 7,197 (3 x 2,399)&lt;/td&gt;
&lt;td&gt;Central billing, admin controls, shared projects, Claude Meetings&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Pro x3, annual billing&lt;/td&gt;
&lt;td&gt;Rs 6,000 (3 x 2,000)&lt;/td&gt;
&lt;td&gt;Individual accounts, Whiteboards, Sheets, Data, no Meetings&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Pro x3, monthly billing&lt;/td&gt;
&lt;td&gt;Rs 7,197 (3 x 2,399)&lt;/td&gt;
&lt;td&gt;Same as above, no annual commitment&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The gap is Rs 1,197 a month, or Rs 14,364 over a year (&lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;). If you would pay Pro on monthly billing anyway, the price is identical and Team is simply the better product. Team monthly pricing is not confirmed in the sources I checked, so verify it on Anthropic's pricing page before you commit.&lt;/p&gt;

&lt;h2&gt;
  
  
  What do Whiteboards, Sheets and Data actually do?
&lt;/h2&gt;

&lt;p&gt;Claude Sheets builds a spreadsheet from a prompt, including tabs, formulas, formatting and drop-downs. When Claude edits a cell, the edit is outlined so you can see what changed. You can share the sheet with colleagues and export to .xlsx or CSV, or continue in Airtable, Asana, monday.com or Smartsheet.&lt;/p&gt;

&lt;p&gt;Claude Whiteboards is a shared brainstorming board. Claude groups sticky notes by theme, drafts action items and can turn the board into a document or a deck. Exports point toward Lucid, Miro and Vani by Zoho.&lt;/p&gt;

&lt;p&gt;Claude Data lets a team run certified queries, marked with a badge, and produce data stories and narrated videos. Partners include Amplitude, Mixpanel, Omni, PostHog and Sigma. For a clinic or studio without an analyst, the practical value is asking plain-language questions of data that someone has already approved.&lt;/p&gt;

&lt;p&gt;On Team plans all three tools are on by default. On Enterprise, Sheets and Whiteboards are off by default and Data switches on 19 November 2026.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you set up Claude Meetings for a small team?
&lt;/h2&gt;

&lt;p&gt;This is a workable sequence for a WhatsApp-first business where clients still arrive by message and calls happen on Google Meet. For the storefront side of that stack, see our &lt;a href="https://dev.to/articles/bluehost-ai-builder-vs-lovable-booking-app"&gt;Bluehost AI Builder vs Lovable booking-app comparison&lt;/a&gt;.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Confirm you are on a Team plan. Claude Meetings is not offered on Pro, so check the plan under organisation settings.&lt;/li&gt;
&lt;li&gt;Enable Meetings. An admin turns the beta on for the workspace.&lt;/li&gt;
&lt;li&gt;Add Claude to Google Meet. It appears as a visible participant, so everyone on the call can see it is there.&lt;/li&gt;
&lt;li&gt;Connect your tools. The interactive connectors open Adobe, Atlassian, Box and Notion as full apps beside the chat. Connect only what the team uses weekly.&lt;/li&gt;
&lt;li&gt;Write handoff rules. Decide who owns the output of every call, for example: Claude drafts action items, the account owner confirms them, and the confirmed list goes to the client on WhatsApp.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Two defaults matter. Claude stays out of calls with outside guests unless an admin allows it, which affects client calls directly. And Claude saves the transcript, summary and screen descriptions, not the audio.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does Claude Meetings cost during the beta?
&lt;/h2&gt;

&lt;p&gt;Notes and transcripts are free on Team and Enterprise during the beta. Answers given in a call are free unless they use connected tools, in which case they count toward plan usage. Documents, whiteboards and artifacts created in a call are free through 10 February 2027, per the &lt;a href="https://claude.com/resources/articles/claude-meetings" rel="noopener noreferrer"&gt;Claude Meetings page&lt;/a&gt;. After that date, check the current terms rather than assuming the same arrangement.&lt;/p&gt;

&lt;p&gt;The beta is not available to organisations on zero data retention, HIPAA-ready setups or customer-managed keys. A small business is unlikely to be in those groups, but a clinic handling patient discussions should ask its advisers before recording consultations.&lt;/p&gt;

&lt;h2&gt;
  
  
  When is Claude Pro x3 still the better buy?
&lt;/h2&gt;

&lt;p&gt;Stay on three Pro seats when each person uses Claude alone, when nobody runs recurring calls, or when you want the lowest bill on annual billing. Pro already includes Sheets, Whiteboards and Data, so you lose only Meetings and central administration. If a business has an owner who wants one invoice and the ability to remove a departed employee's access, that administration is worth the extra Rs 1,197 a month. If you also compare against OpenAI's offering, our piece on the &lt;a href="https://dev.to/articles/openai-business-hub-for-small-business-india"&gt;OpenAI business hub for small businesses in India&lt;/a&gt; covers that side.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: What is the Claude Team plan price in India?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Team Standard is Rs 2,399 per seat per month on annual billing (&lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;). TechCrunch lists the US range as $25 to $30 per seat (&lt;a href="https://techcrunch.com/2026/11/10/anthropic-launches-claude-meetings-to-turn-workplace-discussions-into-plans-and-tasks" rel="noopener noreferrer"&gt;TechCrunch&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How much do three Claude Team seats cost?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Three seats at Rs 2,399 come to Rs 7,197 a month on annual billing (&lt;a href="https://economictimes.indiatimes.com/tech/artificial-intelligence/anthropic-rolls-out-rupee-pricing-for-claude-ai-in-india" rel="noopener noreferrer"&gt;Economic Times&lt;/a&gt;); three Pro plans on annual billing come to Rs 6,000 (&lt;a href="https://moneycontrol.com/technology/claude-india-pricing-for-pro-max-and-team-subscriptions-revealed-here-s-how-much-anthropic-s-ai-plans-cost-article-13972503.html" rel="noopener noreferrer"&gt;Moneycontrol&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Can Claude Pro users use Claude Meetings?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. Claude Meetings is in beta on Team and Enterprise plans only, with Google Meet supported first.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does Claude record the audio of my calls?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. It saves a transcript, a summary and descriptions of shared screens, but not the audio.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Will Claude join calls with clients?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Not by default. It stays out of calls with outside guests unless an admin changes the setting.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Are Sheets and Whiteboards finished products?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. Both launched in beta on 10 November 2026, so expect changes. Claude Data is the only one of the three that is generally available.&lt;/p&gt;

&lt;h2&gt;
  
  
  Note on figures
&lt;/h2&gt;

&lt;p&gt;For context on how this keyword was chosen: we priced 695 keywords in the AI and developer-tooling space using &lt;a href="https://dataforseo.com" rel="noopener noreferrer"&gt;DataForSEO&lt;/a&gt; volume and difficulty data (n=695, measured 2026-11-12), and only 72 (10.4%) cleared a winnable bar of 150-6,000 monthly searches, difficulty 20 or below, a genuine technical term, and at least three words.&lt;/p&gt;

&lt;p&gt;Prices and dates come from the sources linked above. The search-demand estimate of about 300 searches a month for "claude team plan price" comes from a DataForSEO keyword sweep run on 5 November 2026 alongside our pricing articles. It is an estimate, and it was not re-queried for this piece.&lt;/p&gt;

</description>
      <category>claude</category>
      <category>anthropic</category>
      <category>claudeteam</category>
      <category>indiapricing</category>
    </item>
    <item>
      <title>AlphaGenome 2 vs AlphaGenome 1: Cell-Type DNA Effects</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Wed, 18 Nov 2026 04:44:35 +0000</pubDate>
      <link>https://dev.to/aitecharchive/alphagenome-2-vs-alphagenome-1-cell-type-dna-effects-2djn</link>
      <guid>https://dev.to/aitecharchive/alphagenome-2-vs-alphagenome-1-cell-type-dna-effects-2djn</guid>
      <description>&lt;p&gt;AlphaGenome 2 is the better tool for any question about which cell type a DNA variant acts in, and AlphaGenome 1 is now mainly a reference point. On 11 November 2026, Google DeepMind &lt;a href="https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/" rel="noopener noreferrer"&gt;announced the model&lt;/a&gt;, which takes a DNA sequence plus a chosen cell type, tissue or cell state and predicts what a variant does there. On TenK10K, a single-cell map of 28 primary immune cell types, DeepMind reports that it identified 30% of causal variants against 16% for the original model fine-tuned on single-cell data, and 43% of directional gene effects against 23%.&lt;/p&gt;

&lt;p&gt;Last verified: 13 November 2026. The model is for research use only and is not validated or approved for clinical or diagnostic purposes.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;AlphaGenome 2 is a conditional model: one 1-million-letter DNA sequence plus one condition (tissue, cell type or cell state) per prediction.&lt;/li&gt;
&lt;li&gt;On TenK10K, DeepMind reports 30% vs 16% causal variants (&lt;a href="https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/" rel="noopener noreferrer"&gt;DeepMind announcement&lt;/a&gt;) and 43% vs 23% directional gene effects (&lt;a href="https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf" rel="noopener noreferrer"&gt;preprint&lt;/a&gt;), against the original model fine-tuned on single-cell data and other specialised tools.&lt;/li&gt;
&lt;li&gt;It adds DNA methylation as an output and can impute unmeasured properties, such as chromatin accessibility, from one measurement.&lt;/li&gt;
&lt;li&gt;Access is through the AlphaGenome API for non-commercial research, with Colab tutorials; commercial research is in private preview.&lt;/li&gt;
&lt;li&gt;The technical claims come from a preprint that has not been peer reviewed, and the benchmark was run by the model's developers.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What is AlphaGenome 2 and what changed from the first model?
&lt;/h2&gt;

&lt;p&gt;AlphaGenome 2 is a DNA-sequence model built to predict how sequences and their variations affect specific human cell types, with the stated aim of finding the cellular drivers of disease. The first AlphaGenome, launched in 2025, read a 1-million-letter sequence and predicted properties across thousands of experiments from many tissues in parallel. That gave a broad picture but blurred the cell type in which a variant matters.&lt;/p&gt;

&lt;p&gt;The new version changes the architecture rather than just the scale. It is conditional: you supply the sequence and one condition at a time, and that condition modulates the sequence representations at every stage of the network. The question moves from "what does this variant do somewhere in the body" to "what does it do in this cell type, in this state".&lt;/p&gt;

&lt;p&gt;This also addresses a gap in the AlphaGenome Atlas released about two months earlier, in September 2026. The Atlas carried predictions for all 9 billion possible single-letter DNA changes, but, as &lt;a href="https://www.nature.com/articles/d41586-026-03673-0" rel="noopener noreferrer"&gt;Nature noted&lt;/a&gt;, those predictions did not specify the cell type in which a variant acts.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does AlphaGenome 2 compare with AlphaGenome 1?
&lt;/h2&gt;

&lt;p&gt;The table summarises what DeepMind and Nature report. Where the announcement does not state a figure for the first model, the cell says so rather than guessing.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Feature&lt;/th&gt;
&lt;th&gt;AlphaGenome 1&lt;/th&gt;
&lt;th&gt;AlphaGenome 2&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Conditioning&lt;/td&gt;
&lt;td&gt;None; predicts many tissues in parallel&lt;/td&gt;
&lt;td&gt;One condition per run: tissue, cell type or cell state&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Training data&lt;/td&gt;
&lt;td&gt;Not restated in the announcement&lt;/td&gt;
&lt;td&gt;Bulk and single-cell data, human and mouse; over 107 million cell profiles across 68 tissues&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Outputs&lt;/td&gt;
&lt;td&gt;Gene expression, histone modifications, chromatin accessibility&lt;/td&gt;
&lt;td&gt;Same, plus DNA methylation&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TenK10K causal variants&lt;/td&gt;
&lt;td&gt;16% (fine-tuned on single-cell data)&lt;/td&gt;
&lt;td&gt;30%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;TenK10K directional gene effects&lt;/td&gt;
&lt;td&gt;23%&lt;/td&gt;
&lt;td&gt;43%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Access&lt;/td&gt;
&lt;td&gt;See the original release&lt;/td&gt;
&lt;td&gt;API for non-commercial research; commercial in private preview&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Nature reports the model was trained on 637 human and 311 mouse cell types, though users can supply new single-cell gene-activity data to get predictions for states outside that catalogue. The &lt;a href="https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf" rel="noopener noreferrer"&gt;preprint abstract&lt;/a&gt; adds that the model doubles the recall of causal single-cell eQTLs at 90% precision (&lt;a href="https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf" rel="noopener noreferrer"&gt;preprint&lt;/a&gt;) and generated an atlas of enhancer-gene links and scores for 7 million common variants (&lt;a href="https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf" rel="noopener noreferrer"&gt;preprint&lt;/a&gt;) across hundreds of human cell types. Treat those as preprint claims until peer review.&lt;/p&gt;

&lt;h2&gt;
  
  
  Does AlphaGenome 2 reproduce known disease biology?
&lt;/h2&gt;

&lt;p&gt;Yes, in at least two examples DeepMind describes. The first is a 2024 &lt;a href="https://www.nature.com/articles/s41586-024-07501-1" rel="noopener noreferrer"&gt;Nature study&lt;/a&gt; showing that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut macrophages. AlphaGenome 2 singled out the causal variant and showed it selectively boosted ETS2 in gut macrophages, largely leaving other cell types and genes unaffected.&lt;/p&gt;

&lt;p&gt;The second uses unseen single-cell data from a COVID-19 mRNA vaccine study. The model separated vaccine-activated T cells from resting ones and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect missed in unactivated cells. Both are retrospective checks against known or newly supplied data, not prospective discoveries, and the first is a reproduction of a published result.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you use AlphaGenome 2 on your own cell data?
&lt;/h2&gt;

&lt;p&gt;DeepMind describes a workflow that needs no retraining. A researcher uploads a single-cell RNA-seq snapshot, which can be as small as a few dozen cells. The data is compressed into a cell-state representation that conditions the model on a state it never saw in training. DeepMind reports this can match or outperform models trained or fine-tuned directly on those datasets, which is a vendor claim worth testing on your own data.&lt;/p&gt;

&lt;p&gt;The same mechanism supports cross-modality imputation. Given one measurement, such as single-cell gene activity, the model can predict properties you did not measure, such as chromatin accessibility. In practice that helps decide which experiment to run next.&lt;/p&gt;

&lt;h3&gt;
  
  
  How to get started
&lt;/h3&gt;

&lt;ol&gt;
&lt;li&gt;Request access to the AlphaGenome API through DeepMind's announcement page; non-commercial research use is available now.&lt;/li&gt;
&lt;li&gt;Work through the interactive Colab tutorials on expression conditioning to see the input format.&lt;/li&gt;
&lt;li&gt;Pick a condition: a catalogued cell type or tissue, or a cell-state representation built from your own single-cell RNA-seq data.&lt;/li&gt;
&lt;li&gt;Score variants of interest and check the predicted direction (up or down) where the model is confident.&lt;/li&gt;
&lt;li&gt;Use precomputed scores for trait-associated variants, covering traits from height to type 2 diabetes, as a sanity check before running your own.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;DeepMind also says AlphaGenome 2 Skills are coming to Google Antigravity, and it asked researchers to submit use cases by 30 November 2026 for a chance at extra quota and credits.&lt;/p&gt;

&lt;h2&gt;
  
  
  What are the limits of AlphaGenome 2?
&lt;/h2&gt;

&lt;p&gt;There are four to keep in mind. First, the headline numbers are from DeepMind's own preprint, which has not been peer reviewed, so independent replication matters. Second, the 30% and 16% figures describe one benchmark on 28 immune cell types, and performance on other tissues may differ. Third, the model is research only and carries no clinical or diagnostic validation, so it must not inform patient decisions. Fourth, a computational prediction of a causal variant still needs experimental confirmation.&lt;/p&gt;

&lt;p&gt;Independent reaction has been positive. In Nature's coverage, Kyle Lee of the Francis Crick Institute said cell-type specificity was the missing piece that had kept earlier versions from being useful for his work.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is the verdict for genomics and AI-for-science watchers?
&lt;/h2&gt;

&lt;p&gt;For variant interpretation, AlphaGenome 2 wins. Anyone working on immune, gut or other cell-type-specific disease questions gets a model that answers the question they actually ask, and nearly doubling causal-variant recovery on TenK10K is a meaningful step if it replicates. AlphaGenome 1 remains reasonable for broad tissue-level scans, but there is little reason to start new cell-type work on it.&lt;/p&gt;

&lt;p&gt;For wider AI-for-science watchers, the pattern is conditioning a large model on context rather than training one model per dataset. It sits alongside other recent work we have covered, including &lt;a href="https://dev.to/articles/hydrogen-1-0-vs-gpt-6-astra-science-benchmarks-2026"&gt;how GPT-6 Astra and Hydrogen 1.0 compare on science benchmarks&lt;/a&gt;, &lt;a href="https://dev.to/articles/biobasicsbench-gpt-6-astra-wet-lab-benchmark-2026"&gt;BioBasicsBench for wet-lab reasoning&lt;/a&gt;, &lt;a href="https://dev.to/articles/omnii-1-offline-ai-pathogen-screening-laptop"&gt;an offline pathogen-screening model that runs on a laptop&lt;/a&gt;, and &lt;a href="https://dev.to/articles/vault-particle-structure-bacterial-vs-human-claude"&gt;a comparison of bacterial and human vault particle structures&lt;/a&gt;. DeepMind's wider research agenda also appears in &lt;a href="https://dev.to/articles/deepmind-ai-agents-weaken-1024-bit-rsa"&gt;its work on AI agents and 1024-bit RSA&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: What is AlphaGenome 2?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It is a DeepMind DNA-sequence model that predicts how DNA sequences and variants affect specific human cell types, using a chosen cell type, tissue or cell state as an input condition.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How much better is AlphaGenome 2 than the original model?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; On the TenK10K single-cell benchmark, DeepMind reports 30% of causal variants identified against 16%, and 43% of directional gene effects against 23%, compared with the original model fine-tuned on single-cell data. These figures come from a preprint.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Can I use AlphaGenome 2 for clinical or diagnostic decisions?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. DeepMind states it is for research use only and is not validated or approved for clinical or diagnostic purposes.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How do I access AlphaGenome 2?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Non-commercial research use is available through the AlphaGenome API, with Colab tutorials on expression conditioning. Commercial research is in private preview.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does AlphaGenome 2 need retraining for new cell types?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. You can upload a single-cell RNA-seq snapshot, even a few dozen cells, and the model converts it into a cell-state representation used to condition predictions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: What new outputs does AlphaGenome 2 add?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It adds DNA methylation to the existing outputs of gene expression, histone modifications and chromatin accessibility, and can impute unmeasured properties from a single measurement.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/" rel="noopener noreferrer"&gt;DeepMind announcement&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf" rel="noopener noreferrer"&gt;AlphaGenome 2 preprint (not peer reviewed)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.nature.com/articles/d41586-026-03673-0" rel="noopener noreferrer"&gt;Nature news coverage&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://the-decoder.com/google-deepmind-shifts-genomics-ai-from-bulk-tissue-predictions-to-single-cell-type-resolution-with-alphagenome-2" rel="noopener noreferrer"&gt;The Decoder report&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://www.nature.com/articles/s41586-024-07501-1" rel="noopener noreferrer"&gt;ETS2 inflammatory bowel disease study, Nature 2024&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Corrections log: none at publication. This article was produced with AI assistance and checked against the primary sources above.&lt;/p&gt;

</description>
      <category>alphagenome2</category>
      <category>deepmind</category>
      <category>genomics</category>
      <category>singlecell</category>
    </item>
    <item>
      <title>India AI Satellite Constellation Gets Rs 64.05 Crore</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Wed, 18 Nov 2026 04:44:15 +0000</pubDate>
      <link>https://dev.to/aitecharchive/india-ai-satellite-constellation-gets-rs-6405-crore-956</link>
      <guid>https://dev.to/aitecharchive/india-ai-satellite-constellation-gets-rs-6405-crore-956</guid>
      <description>&lt;p&gt;The Technology Development Board (TDB), part of the Department of Science and Technology (DST), signed an agreement with Astrome Technologies Pvt Ltd of Bengaluru on 12 November 2026. It commits Rs 64.05 crore from the Research, Development and Innovation (RDI) Fund towards a project costing Rs 128.11 crore, according to the &lt;a href="https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2335322&amp;amp;reg=48&amp;amp;lang=2" rel="noopener noreferrer"&gt;PIB release&lt;/a&gt;. The goal is a sovereign India AI satellite constellation for electronic intelligence (AI-ELINT), built from 50 kg microsatellites in low Earth orbit.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;By Sham Uddin. Last verified: 13 November 2026.&lt;/em&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;TDB is funding 50 percent of a Rs 128.11 crore project, per the &lt;a href="https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2335322&amp;amp;reg=48&amp;amp;lang=2" rel="noopener noreferrer"&gt;PIB release&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;The money is Optionally Convertible Debt (OCD), not a grant.&lt;/li&gt;
&lt;li&gt;The work takes the technology from Technology Readiness Level (TRL) 4 to TRL 9.&lt;/li&gt;
&lt;li&gt;Three pieces are being built: an ELINT payload, AI ground processing software, and inter-satellite links.&lt;/li&gt;
&lt;li&gt;TDB's project sheet describes launching 3 satellites first, with a much larger constellation as the long-term plan.&lt;/li&gt;
&lt;li&gt;The programme depends as much on signal-triage software as on the satellites.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What exactly did TDB and Astrome sign?
&lt;/h2&gt;

&lt;p&gt;The PIB release states that the support comes from the RDI Fund and equals half of the total outlay of Rs 128.11 crore (&lt;a href="https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2335322&amp;amp;reg=48&amp;amp;lang=2" rel="noopener noreferrer"&gt;PIB&lt;/a&gt;), does not detail who funds the balance, and places the project in the Deep Technology sector, Space Technology sub-sector, for national security applications.&lt;/p&gt;

&lt;p&gt;TDB Secretary Rajesh Kumar Pathak framed it as support for Indian enterprises working on frontier technologies that strengthen strategic capability. An Astrome representative said the funding would speed up integration of its ELINT payloads, AI data processing and inter-satellite links. &lt;a href="https://www.thehansindia.com/business/govt-signs-rs-64-crore-deal-with-astrome-for-ai-enabled-intelligence-satellite-constellation-1141399" rel="noopener noreferrer"&gt;Hans India via IANS&lt;/a&gt; reported the same figures.&lt;/p&gt;

&lt;h2&gt;
  
  
  What is AI-ELINT and why does it need AI?
&lt;/h2&gt;

&lt;p&gt;Electronic intelligence, or ELINT, means collecting and analysing radio and radar emissions rather than taking pictures. Optical satellites are limited by cloud and darkness; ELINT listens regardless of weather or time of day, a contrast also drawn in &lt;a href="https://www.news18.com/science/indias-rs-64-crore-space-spy-push-centre-backs-indigenous-ai-powered-satellite-network-ws-l-10423770.html" rel="noopener noreferrer"&gt;News18's coverage&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;The AI part is a data-handling problem. TDB's &lt;a href="https://tdb.gov.in/sites/default/files/2026-06/rdi_scheme.pdf" rel="noopener noreferrer"&gt;RDI scheme document&lt;/a&gt; describes a 70 MHz to 18 GHz payload, sensitive enough to pick up anything from satellite phones to missile telemetry. A receiver that wide produces far more raw signal than a small satellite can send to the ground. That is our reading of the design problem, not a project figure, but it explains why the AI layer is a core deliverable.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Stage&lt;/th&gt;
&lt;th&gt;Without automation&lt;/th&gt;
&lt;th&gt;With AI-based processing&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Capture&lt;/td&gt;
&lt;td&gt;Wideband samples recorded on the satellite&lt;/td&gt;
&lt;td&gt;Same capture, with detection applied early&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Downlink&lt;/td&gt;
&lt;td&gt;Limited by the satellite's bandwidth and pass time&lt;/td&gt;
&lt;td&gt;Prioritised emitters and summaries sent first&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Analysis&lt;/td&gt;
&lt;td&gt;Analysts search raw spectrum manually&lt;/td&gt;
&lt;td&gt;Emitters are classified and geolocated automatically&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Output&lt;/td&gt;
&lt;td&gt;Large backlog of unreviewed data&lt;/td&gt;
&lt;td&gt;Shorter queue of ranked signals of interest&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The table is an analytical illustration, not a published specification of Astrome's system.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does the India AI satellite constellation actually consist of?
&lt;/h2&gt;

&lt;p&gt;The PIB release lists three components: an ELINT payload that detects and analyses signals over a broad frequency range, AI-based ground software for processing constellation data, and inter-satellite links that let satellites work as one network (&lt;a href="https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2335322&amp;amp;reg=48&amp;amp;lang=2" rel="noopener noreferrer"&gt;PIB&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;Scale needs careful reading. The TDB project sheet gives the near-term goal as 3 AI-ELINT microsatellites of the 50 kg class, on an inclined orbit at 40 degrees about 600 km up, with eventual progressive deployment of 231 satellites (&lt;a href="https://tdb.gov.in/sites/default/files/2026-06/rdi_scheme.pdf" rel="noopener noreferrer"&gt;TDB&lt;/a&gt;). The Rs 64.05 crore agreement should therefore be read as funding for the first technology-proving stage, not the full 231-satellite build.&lt;/p&gt;

&lt;p&gt;Astrome is a plausible fit for the links component: its &lt;a href="https://astrome.co/" rel="noopener noreferrer"&gt;company site&lt;/a&gt; says it was founded in 2015 and lists phased-array antennas, software-defined modems and satellite products, including an inter-satellite link network.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does moving from TRL 4 to TRL 9 involve?
&lt;/h2&gt;

&lt;p&gt;TRL is a nine-point maturity scale. The PIB release says the programme advances critical technologies from TRL 4 to TRL 9 through development, engineering, integration, testing and validation. Under the standard scale described by &lt;a href="https://www.nasa.gov/directorates/somd/space-communications-navigation-program/technology-readiness-levels/" rel="noopener noreferrer"&gt;NASA&lt;/a&gt;:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;TRL&lt;/th&gt;
&lt;th&gt;General meaning&lt;/th&gt;
&lt;th&gt;What it implies here&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;4&lt;/td&gt;
&lt;td&gt;Components validated in a laboratory&lt;/td&gt;
&lt;td&gt;Payload and processing code work on a bench&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;5&lt;/td&gt;
&lt;td&gt;Components validated in a relevant environment&lt;/td&gt;
&lt;td&gt;Hardware tested under space-like conditions&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;6&lt;/td&gt;
&lt;td&gt;Prototype demonstrated in a relevant environment&lt;/td&gt;
&lt;td&gt;Integrated payload and software tested together&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;Prototype demonstrated in space&lt;/td&gt;
&lt;td&gt;First satellite operating in orbit&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;8&lt;/td&gt;
&lt;td&gt;System complete and qualified&lt;/td&gt;
&lt;td&gt;Flight design proven against requirements&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;9&lt;/td&gt;
&lt;td&gt;System proven in operations&lt;/td&gt;
&lt;td&gt;Satellites performing the mission reliably&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;No dates for these milestones appear in the sources reviewed; the rightmost column is our interpretation of the general scale.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does Optionally Convertible Debt mean for the startup?
&lt;/h2&gt;

&lt;p&gt;The PIB release names OCD as the funding mode: in general terms, a loan the lender can later convert into equity. For Astrome, the Rs 64.05 crore is not free money but also not an immediate sale of ownership. For the state, it means a stake in the upside if the technology works, while the company carries repayment obligations if it does not. The precise conversion terms were not published in the sources reviewed.&lt;/p&gt;

&lt;p&gt;This is also a shift in who builds strategic space assets: programmes of this kind were historically the domain of state laboratories or foreign suppliers, while the RDI scheme funds private deep-tech firms to do the work.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where does this sit in the wider RDI Fund?
&lt;/h2&gt;

&lt;p&gt;The RDI Scheme was approved on 1 July 2025 with Rs 1 lakh crore over six years; TDB acts as second-level fund manager with an initial Rs 2,000 crore allocation (&lt;a href="https://tdb.gov.in/sites/default/files/2026-06/rdi_scheme.pdf" rel="noopener noreferrer"&gt;TDB&lt;/a&gt;). As of 30 April 2026 it records Rs 4,743.99 crore in approved project cost, Rs 2,192.32 crore of approved TDB assistance, and 22 companies selected.&lt;/p&gt;

&lt;p&gt;In October 2026, TDB agreed Rs 116 crore for Manastu Space Technologies against a Rs 277.63 crore green-propulsion project, as &lt;a href="https://www.thehansindia.com/business/govt-signs-rs-64-crore-deal-with-astrome-for-ai-enabled-intelligence-satellite-constellation-1141399" rel="noopener noreferrer"&gt;Hans India reported&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;For other threads in India's AI push, see our coverage of the &lt;a href="https://dev.to/articles/indiaai-foundation-models-eka-1-dhir-2b-dhir-5b"&gt;IndiaAI foundation models EKA-1 and Dhir&lt;/a&gt;, the &lt;a href="https://dev.to/articles/andhra-reliance-ai-data-centre-bhogapuram-25-acres"&gt;Andhra Pradesh AI data centre at Bhogapuram&lt;/a&gt;, &lt;a href="https://dev.to/articles/indian-railways-ai-cameras-700-festive-war-room-2026"&gt;Indian Railways' AI cameras&lt;/a&gt;, the &lt;a href="https://dev.to/articles/npci-tech-solutions-ntsl-subsidiary-ai-blockchain"&gt;NPCI technology subsidiary&lt;/a&gt;, and a &lt;a href="https://dev.to/articles/hindi-text-to-speech-api-chariot-1-tts-multilingual-launch"&gt;Hindi text-to-speech API launch&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  What are the open questions?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Launch dates and providers are not stated in the release.&lt;/li&gt;
&lt;li&gt;Astrome's own contribution beyond the 50 percent split, and the OCD terms, are not public.&lt;/li&gt;
&lt;li&gt;The step from 3 satellites to 231 has no published funding or timeline.&lt;/li&gt;
&lt;li&gt;Performance claims come from TDB's project sheet, not independent testing.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: How much is the TDB-Astrome agreement worth?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; TDB is providing Rs 64.05 crore from the RDI Fund against a Rs 128.11 crore total project cost, 50 percent of the outlay (&lt;a href="https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2335322&amp;amp;reg=48&amp;amp;lang=2" rel="noopener noreferrer"&gt;PIB&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: What is an AI-ELINT satellite?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; A satellite that collects radio and radar emissions and uses AI to detect, classify and prioritise them, so analysts get ranked signals rather than raw data.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How many satellites will be built?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; TDB's project sheet lists 3 microsatellites of the 50 kg class as the immediate goal, with 231 as the eventual deployment (&lt;a href="https://tdb.gov.in/sites/default/files/2026-06/rdi_scheme.pdf" rel="noopener noreferrer"&gt;TDB&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is the Rs 64.05 crore a grant?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. The PIB release says the support is Optionally Convertible Debt, a loan that can convert into equity (&lt;a href="https://www.pib.gov.in/PressReleseDetailm.aspx?PRID=2335322&amp;amp;reg=48&amp;amp;lang=2" rel="noopener noreferrer"&gt;PIB&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Who is Astrome Technologies?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; A Bengaluru deep-tech company founded in 2015 that works on wireless and satellite communications, including phased-array antennas and inter-satellite links (&lt;a href="https://astrome.co/" rel="noopener noreferrer"&gt;Astrome&lt;/a&gt;).&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This article was prepared with AI assistance and edited for accuracy. Figures come from the linked primary sources, and corrections are welcome.&lt;/em&gt;&lt;/p&gt;

</description>
      <category>indiaaisatelliteconstellation</category>
      <category>aielint</category>
      <category>astrometechnologies</category>
      <category>technologydevelopmentboard</category>
    </item>
    <item>
      <title>Claude Code 2.1.323: Code Mode vs Standard Tool Calls</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Wed, 18 Nov 2026 04:43:55 +0000</pubDate>
      <link>https://dev.to/aitecharchive/claude-code-21323-code-mode-vs-standard-tool-calls-2al0</link>
      <guid>https://dev.to/aitecharchive/claude-code-21323-code-mode-vs-standard-tool-calls-2al0</guid>
      <description>&lt;p&gt;Claude Code 2.1.323, released on 12 November 2026, adds an experimental, opt-in Code Mode in which Claude writes JavaScript to call its tools, and it ships alongside dozens of Bash permission and sandbox fixes. Our verdict: standard tool calls win for day-to-day work today, and the upgrade is worth taking for the hardening alone. Trial Code Mode only in a sandboxed or throwaway project until its documentation is live and your team has decided who may turn it on.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Claude Code 2.1.323 adds Code Mode, an experimental opt-in where Claude writes JavaScript to call its tools. Organisations can turn it off.&lt;/li&gt;
&lt;li&gt;The same release fixes dozens of Bash permission and sandbox problems, several of which let commands run without a prompt.&lt;/li&gt;
&lt;li&gt;Verdict: keep standard tool calls as your default and upgrade for the fixes. Treat Code Mode as a supervised experiment.&lt;/li&gt;
&lt;li&gt;The Code Mode documentation page returned a 404 when we checked on 13 November 2026, so this article does not include setup steps.&lt;/li&gt;
&lt;li&gt;If you use plugins, the new &lt;code&gt;engines["claude-code"]&lt;/code&gt; field in &lt;code&gt;plugin.json&lt;/code&gt; stops plugins loading outside their declared version range.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  What is Code Mode in Claude Code 2.1.323?
&lt;/h2&gt;

&lt;p&gt;The &lt;a href="https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md" rel="noopener noreferrer"&gt;official changelog&lt;/a&gt; describes it as an experimental opt-in in which Claude writes JavaScript to call its tools. The entry points to a docs page for turning it on, and for turning it off across an organisation.&lt;/p&gt;

&lt;p&gt;Under standard tool calling, the model emits one structured request per tool call, waits for the result, then decides what to do next. Code Mode changes the unit of work. Claude writes a short program that calls several tools, and the program handles the intermediate results itself. Amazon Web Services' partner blog &lt;a href="https://dev.classmethod.jp/en/articles/20261113-cc-updates-v2-1-323" rel="noopener noreferrer"&gt;published a walkthrough of the release the same morning&lt;/a&gt; and flagged the same org-level off switch as the notable governance detail.&lt;/p&gt;

&lt;p&gt;A concrete example: finding every file that mentions a deprecated function, then reading the three largest. With standard calls that is a search, a result returned to the model, then three reads, each a round trip. In Code Mode it could be one script that searches, sorts and reads in a single execution.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code Mode vs standard tool calls: which should you use?
&lt;/h2&gt;

&lt;p&gt;The design intent is fewer round trips through the model, which should reduce overhead in long sessions. We have no measurements of that for 2.1.323, and Anthropic has not published any that we could find, so treat it as a hypothesis.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;&lt;/th&gt;
&lt;th&gt;Standard tool calls&lt;/th&gt;
&lt;th&gt;Code Mode&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Status&lt;/td&gt;
&lt;td&gt;Default, mature&lt;/td&gt;
&lt;td&gt;Experimental, opt-in&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;What you review&lt;/td&gt;
&lt;td&gt;One discrete call at a time&lt;/td&gt;
&lt;td&gt;A script that may make many calls&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Organisation control&lt;/td&gt;
&lt;td&gt;Existing permission rules&lt;/td&gt;
&lt;td&gt;Can be turned off organisation-wide&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Documentation&lt;/td&gt;
&lt;td&gt;Available&lt;/td&gt;
&lt;td&gt;Docs page returned 404 on 13 Nov 2026&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Best for&lt;/td&gt;
&lt;td&gt;Day-to-day edits, reviews, anything on a real repository&lt;/td&gt;
&lt;td&gt;Supervised trials in disposable projects&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Standard tool calls win for now because their failure modes are known. A permission prompt shows you one action. A script can bundle several actions into a single approval, so reviewing it takes more care. That is a tradeoff, not a flaw, but it matters when the surrounding permission system has just needed this many fixes.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does Claude Code 2.1.323 change for Bash permissions?
&lt;/h2&gt;

&lt;p&gt;Most of this release is not Code Mode. It is a hardening pass on the places where a prompt-injected instruction could slip through. The changelog lists these among its fixes:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Some &lt;code&gt;lsof&lt;/code&gt; commands ran without a permission prompt.&lt;/li&gt;
&lt;li&gt;The Bash sandbox could switch off during a session.&lt;/li&gt;
&lt;li&gt;Allow rules with an attached wildcard, such as &lt;code&gt;Bash(ls*)&lt;/code&gt;, skipped the prompt in some cases. A rule that names a variable, such as &lt;code&gt;Bash(NODE_ENV*)&lt;/code&gt;, now asks.&lt;/li&gt;
&lt;li&gt;Permission checks could judge an &lt;code&gt;rm&lt;/code&gt; from the wrong folder and skip the dangerous-removal prompt under bypass permissions or a Bash allow rule.&lt;/li&gt;
&lt;li&gt;Sandboxed commands could read Claude Code's file checkpoint copies under the config directory.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;claude -p&lt;/code&gt; and the Agent SDK ran Bash unsandboxed when &lt;code&gt;--settings&lt;/code&gt; or SDK settings that enable the sandbox failed validation. Claude Code now stops at start-up and names the invalid setting.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A further fix concerns auto mode. If &lt;code&gt;--permission-mode auto&lt;/code&gt; or &lt;code&gt;defaultMode: auto&lt;/code&gt; is set but auto mode is unavailable, Claude Code previously started in a mode that could allow more than intended. It now starts in Manual mode, which asks first. The &lt;a href="https://daily.rclabs.in/i/claude-code-v2-1-323-8427d4" rel="noopener noreferrer"&gt;release notes index&lt;/a&gt; confirms the same security batch, including the wildcard allow-rule fix and the new engine identifier for plugins.&lt;/p&gt;

&lt;p&gt;The wildcard fix is the one to act on. If your settings contain allow rules such as &lt;code&gt;Bash(ls*)&lt;/code&gt;, review them after upgrading. Some will now prompt where they previously did not, and that is the correct behaviour.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does this release treat plugins?
&lt;/h2&gt;

&lt;p&gt;Three changes affect plugin authors and anyone installing third-party plugins:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;plugin.json&lt;/code&gt; accepts &lt;code&gt;engines["claude-code"]&lt;/code&gt;. Plugins outside their stated Claude Code version range are not installed, updated or loaded, and install and update choose the newest version that fits.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;claude plugin validate&lt;/code&gt; now type-checks a plugin's TypeScript hooks module. Each finding is a warning, and &lt;code&gt;--strict&lt;/code&gt; makes it fail.&lt;/li&gt;
&lt;li&gt;Mods can no longer reword Claude Code's own confirmation questions.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The last item is worth noting for security. A confirmation prompt is only useful if the text is trustworthy, so stopping extensions from rewriting it removes a real social-engineering route.&lt;/p&gt;

&lt;h2&gt;
  
  
  Should you upgrade to Claude Code 2.1.323?
&lt;/h2&gt;

&lt;p&gt;Yes, for the fixes. The 2.1.323 release arrived two days after 2.1.321 and three days after 2.1.322, and the &lt;a href="https://havoptic.com/r/claude-code-2.1.323" rel="noopener noreferrer"&gt;Havoptic release tracker&lt;/a&gt; counts it as the 262nd Claude Code release of 2026. Independent version trackers &lt;a href="https://claudeupdates.dev/version/2.1.323" rel="noopener noreferrer"&gt;mirror the same changelog batch&lt;/a&gt;. At that cadence, pinning a version for months is hard to justify when the notes describe prompts being skipped.&lt;/p&gt;

&lt;p&gt;A sensible sequence:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Upgrade a single machine and run your usual workflow with standard tool calls.&lt;/li&gt;
&lt;li&gt;Review any Bash allow rules ending in a wildcard.&lt;/li&gt;
&lt;li&gt;If you use &lt;code&gt;claude -p&lt;/code&gt; or the Agent SDK in CI, confirm your sandbox settings validate, since an invalid setting now stops start-up instead of running unsandboxed.&lt;/li&gt;
&lt;li&gt;Decide at organisation level whether Code Mode should be available at all, before anyone enables it individually.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Secrets handling matters more once a model writes executable code. Our guide to &lt;a href="https://dev.to/articles/bitwarden-vs-1password-claude-code-secrets-security-setup"&gt;Bitwarden vs 1Password for Claude Code secrets&lt;/a&gt; covers keeping credentials out of reach. For the previous release's default-model and pricing changes, see our &lt;a href="https://dev.to/articles/claude-code-2-1-321-fable-5-2-default-price-security"&gt;Claude Code 2.1.321 coverage&lt;/a&gt;. If you are weighing how much to trust agent output, &lt;a href="https://dev.to/articles/agentic-engineering-verify-first-vs-build-first"&gt;verify-first vs build-first&lt;/a&gt; is a useful frame, and our piece on &lt;a href="https://dev.to/articles/claude-code-subagents-99-percent-cheaper-status-updates"&gt;subagent cost and status updates&lt;/a&gt; covers the spending side.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does it cost, including in India?
&lt;/h2&gt;

&lt;p&gt;Claude Code is included with Claude Pro and Max subscriptions, and Anthropic lists prices on its &lt;a href="https://claude.com/pricing" rel="noopener noreferrer"&gt;pricing page&lt;/a&gt;. Pricing is in US dollars. Indian users generally pay by international card, and GST may apply to imported software services, so check the final charge at checkout. This release does not change the pricing structure.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why we cover this
&lt;/h2&gt;

&lt;p&gt;We choose topics partly by measuring demand. In our own keyword corpus (n=695, measured 12 November 2026), we priced 695 keywords in the AI and developer-tooling space using DataForSEO volume and difficulty data. Only a small share cleared our bar for a winnable topic: modest search volume, low difficulty, a genuine technical term, and a phrase of at least three words. Version-specific release coverage like this is a narrow niche, and the audience for it is people who need to decide what to do this week.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: What is Code Mode in Claude Code 2.1.323?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It is an experimental, opt-in feature in which Claude writes JavaScript to call its tools instead of issuing each tool call separately. The changelog says organisations can turn it off.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How do I enable Code Mode?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The changelog points to a documentation page for this, but that page returned a 404 when we checked on 13 November 2026. We have not verified the steps, so check the official docs before relying on any guide.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is Claude Code 2.1.323 safe to upgrade to?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The release is mostly fixes, including several that close cases where Bash commands ran without a prompt. Upgrade one machine first and review any wildcard allow rules, since some will now ask where they did not before.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does Code Mode make Claude Code cheaper or faster?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The idea is fewer model round trips in long sessions, but we found no published measurements for 2.1.323. Treat any savings as unproven until you test on your own workload.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: What happens if auto mode is unavailable?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Claude Code now starts in Manual mode, which asks before acting. Previously it could start in a mode that allowed more than intended.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Can plugins break after this update?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Plugins that declare an &lt;code&gt;engines["claude-code"]&lt;/code&gt; range are not installed, updated or loaded outside it. Authors should run &lt;code&gt;claude plugin validate&lt;/code&gt;, optionally with &lt;code&gt;--strict&lt;/code&gt;, to catch hooks type errors.&lt;/p&gt;

</description>
      <category>claudecode</category>
      <category>codemode</category>
      <category>bashpermissions</category>
      <category>promptinjection</category>
    </item>
    <item>
      <title>Inkling 2 Small vs Muse Spark Open: Best US Open Model</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Wed, 18 Nov 2026 04:43:34 +0000</pubDate>
      <link>https://dev.to/aitecharchive/inkling-2-small-vs-muse-spark-open-best-us-open-model-7bh</link>
      <guid>https://dev.to/aitecharchive/inkling-2-small-vs-muse-spark-open-best-us-open-model-7bh</guid>
      <description>&lt;p&gt;Inkling 2 Small is the best US open-weight model for data analysis that Plotly has tested, and it beats Meta's Muse Spark Open by one question. Thinking Machines Lab &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;published the weights on November 12, 2026&lt;/a&gt; under Apache 2.0, and Plotly's &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;43-question benchmark&lt;/a&gt; shows it passing 39 (91%) against 38 for Muse Spark Open, at $1.81 per full run against $5.16. If you want an open-weight model from a US lab, Inkling 2 Small is the one to pick. If raw accuracy matters more than origin, Kimi K3.1 Pro still scores higher.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Inkling 2 Small is a &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;594B-parameter Mixture-of-Experts model with 17B active per token&lt;/a&gt;, accepting text and images.&lt;/li&gt;
&lt;li&gt;It passed 39 of 43 questions in &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly's data-analysis test&lt;/a&gt;, third among open-weight models overall.&lt;/li&gt;
&lt;li&gt;Hosted pricing is &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;$0.30 per million input tokens and $1.00 per million output tokens&lt;/a&gt;, or Rs 28 and Rs 94 at an &lt;a href="https://www.divybharat.com/markets/usd-to-inr" rel="noopener noreferrer"&gt;exchange rate of 94.05&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Self-hosting the 4-bit checkpoint needs &lt;a href="https://thinkingmachines.ai/model-card/inkling-2-small/" rel="noopener noreferrer"&gt;at least 500GB of aggregated VRAM&lt;/a&gt;, though Plotly notes it fits in 356GB of memory.&lt;/li&gt;
&lt;li&gt;Weak spots: finding facts buried in large inputs, and a chat style the vendor admits is unpolished.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  How does Inkling 2 Small compare with other open-weight models?
&lt;/h2&gt;

&lt;p&gt;Plotly's benchmark is a data-analysis exam built from real-world tasks. The table below uses its published figures; the &lt;a href="https://www.divybharat.com/markets/usd-to-inr" rel="noopener noreferrer"&gt;price conversion&lt;/a&gt; uses Rs 94.05 per dollar.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Score (of 43)&lt;/th&gt;
&lt;th&gt;Cost per run&lt;/th&gt;
&lt;th&gt;Median time per question&lt;/th&gt;
&lt;th&gt;Open weights&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Kimi K3.1 Pro&lt;/td&gt;
&lt;td&gt;42&lt;/td&gt;
&lt;td&gt;not stated here&lt;/td&gt;
&lt;td&gt;about 2.5 minutes&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;MiniMax M3.1 Flash&lt;/td&gt;
&lt;td&gt;41&lt;/td&gt;
&lt;td&gt;not stated here&lt;/td&gt;
&lt;td&gt;about 3 minutes&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Inkling 2 Small&lt;/td&gt;
&lt;td&gt;39&lt;/td&gt;
&lt;td&gt;$1.81 (Rs 170)&lt;/td&gt;
&lt;td&gt;13 seconds&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Muse Spark Open&lt;/td&gt;
&lt;td&gt;38&lt;/td&gt;
&lt;td&gt;$5.16 (Rs 485)&lt;/td&gt;
&lt;td&gt;48 seconds&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Inkling-Small&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;$1.93 (Rs 182)&lt;/td&gt;
&lt;td&gt;25 seconds&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Particle-1&lt;/td&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;td&gt;$0.35 (Rs 33)&lt;/td&gt;
&lt;td&gt;16 seconds&lt;/td&gt;
&lt;td&gt;No&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Source for all rows: &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly&lt;/a&gt;. Two readings matter. First, within open weights, Plotly says no model is both cheaper and more accurate than Inkling 2 Small, which puts it on the cost-accuracy frontier. Second, across all models the closed Particle-1 scored 40 for $0.35 per run (&lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly&lt;/a&gt;), so Inkling 2 Small is not the overall bargain. Against its predecessor the gain is nine questions for about the same money, and the run takes roughly half the time.&lt;/p&gt;

&lt;p&gt;For a fuller look at the accuracy leader, see our coverage of the &lt;a href="https://dev.to/articles/kimi-k3-1-pro-open-weights-released"&gt;Kimi K3.1 Pro open-weights release&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where does Inkling 2 Small do well, and where does it fail?
&lt;/h2&gt;

&lt;p&gt;Per the &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly results&lt;/a&gt;, it scored 5/5 on impossible requests, 3/3 on data-quality problems and 5/5 on statistics traps. In practice that means it declined questions the data could not answer, caught all three planted data issues, and avoided every statistical trap.&lt;/p&gt;

&lt;p&gt;It lost four questions. One was an analytics question where it wrote its final table call as text instead of running it. The other three were hidden-needle questions, where an answer sits inside a large body of material; it found 2 of 5. Muse Spark Open also found 2, while Kimi K3.1 Pro and MiniMax M3.1 Flash found all five. If your workload is retrieval from long documents, this is the limitation to test first.&lt;/p&gt;

&lt;p&gt;Efficiency helps explain the low bill. Plotly measured about 5,700 output tokens per question, against about 7,800 for Muse Spark Open and 32,000 for MiniMax M3.1 Flash, and about 96% of input tokens came from the provider's cache (see the &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;same report&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;Thinking Machines is open about the other weaknesses. In its &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;announcement&lt;/a&gt;, it says the model is less tuned for chat, with weaker content organisation, overuse of slashes, and overreaction to mental-health and safety-related queries.&lt;/p&gt;

&lt;h2&gt;
  
  
  What does Inkling 2 Small cost?
&lt;/h2&gt;

&lt;p&gt;Hosted access through OpenRouter, with Together and Cloudflare serving, is priced as follows, according to &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly&lt;/a&gt;. Rupee figures use the &lt;a href="https://www.divybharat.com/markets/usd-to-inr" rel="noopener noreferrer"&gt;Rs 94.05 rate&lt;/a&gt; as of November 13, 2026.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Item&lt;/th&gt;
&lt;th&gt;USD&lt;/th&gt;
&lt;th&gt;INR&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Input, per million tokens&lt;/td&gt;
&lt;td&gt;$0.30&lt;/td&gt;
&lt;td&gt;about Rs 28&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Output, per million tokens&lt;/td&gt;
&lt;td&gt;$1.00&lt;/td&gt;
&lt;td&gt;about Rs 94&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Context window&lt;/td&gt;
&lt;td&gt;1M tokens&lt;/td&gt;
&lt;td&gt;not applicable&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;For a price comparison against coding-focused rivals, see &lt;a href="https://dev.to/articles/ripple-65b-vs-open-weight-coding-models-price-2026"&gt;Ripple 65B versus open-weight coding models&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How do you run Inkling 2 Small yourself?
&lt;/h2&gt;

&lt;p&gt;There are four routes, from least to most effort.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Try it hosted. The &lt;a href="https://tinker.thinkingmachines.ai/playground" rel="noopener noreferrer"&gt;Tinker Playground&lt;/a&gt; is free to try, and Tinker also handles fine-tuning and sampling.&lt;/li&gt;
&lt;li&gt;Call a managed API. OpenRouter has a paid endpoint, and the Hugging Face page lists providers including NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.&lt;/li&gt;
&lt;li&gt;Download the weights. The &lt;a href="https://huggingface.co/thinkingmachines/Inkling-2-Small" rel="noopener noreferrer"&gt;NVFP4 checkpoint&lt;/a&gt; and the &lt;a href="https://huggingface.co/thinkingmachines/Inkling-2-Small-MXFP8" rel="noopener noreferrer"&gt;MXFP8 checkpoint&lt;/a&gt; are both Apache 2.0.&lt;/li&gt;
&lt;li&gt;Serve it locally with an official recipe: &lt;a href="https://recipes.vllm.ai/thinkingmachines/Inkling-2-Small" rel="noopener noreferrer"&gt;vLLM&lt;/a&gt; or &lt;a href="https://docs.sglang.io/cookbook/autoregressive/ThinkingMachines/Inkling-2-Small" rel="noopener noreferrer"&gt;SGLang&lt;/a&gt;. Unsloth and TokenSpeed are also supported, and Miles can post-train it.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;On hardware, the &lt;a href="https://thinkingmachines.ai/model-card/inkling-2-small/" rel="noopener noreferrer"&gt;model card&lt;/a&gt; calls for a GPU cluster with at least 500GB of aggregated VRAM for the main checkpoint, with two NVIDIA B300 cards as an example. Plotly says the 4-bit version needs about 356GB, which fits a single 512GB Mac Studio. The MXFP8 build runs on one server node.&lt;/p&gt;

&lt;p&gt;Reported serving figures, from &lt;a href="https://aiaulis.com/en/juttu/thinking-machines-releases-inkling-2-small-weights-under-apache-2-0-cac8a9148f36" rel="noopener noreferrer"&gt;measurements by Thinking Machines and Inferact&lt;/a&gt;: the model has 48 layers and 384 experts per layer with 8 active per token. A single user gets about 630 tokens per second on four GB200 GPUs with speculative decoding. On an agentic workload across 8 GPUs, throughput falls from about 1,070 tokens per second per GPU at 32K context to about 450 at 256K, as concurrent users drop from roughly 84 to 37 (&lt;a href="https://aiaulis.com/en/juttu/thinking-machines-releases-inkling-2-small-weights-under-apache-2-0-cac8a9148f36" rel="noopener noreferrer"&gt;AI Aulis, citing the vLLM blog measurements&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;If MIT-licensed self-hosting is your priority, compare it with &lt;a href="https://dev.to/articles/longcat-2-5-open-weights-mit-self-host"&gt;LongCat 2.5&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does it fare against closed frontier models?
&lt;/h2&gt;

&lt;p&gt;On GPU kernel optimisation, the company's own benchmark shows a gap. On a DeepSeek-V3.2 sparse-attention kernel, Inkling 2 Small reached a &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;6.02x speedup, against 18.80x for Claude Opus 5.5 and 10.26x for GPT-6 Astra&lt;/a&gt;, and these apply to that kernel alone, not end-to-end training. Its &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;494.27 AUC on PMPP-Hard&lt;/a&gt; was the second-best reported in that row. For closed-model context, see our pieces on &lt;a href="https://dev.to/articles/claude-fable-5-2-vs-fable-5-1-price-benchmarks-2026"&gt;Claude Fable 5.2 versus 5.1&lt;/a&gt; and &lt;a href="https://dev.to/articles/hydrogen-1-0-vs-gpt-6-astra-science-benchmarks-2026"&gt;Hydrogen 1.0 versus GPT-6 Astra&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Which model should you pick?
&lt;/h2&gt;

&lt;p&gt;Pick Inkling 2 Small if you need an open-weight model from a US lab, want quick answers at low cost, and your work is analysis rather than needle-in-a-haystack retrieval. Pick Kimi K3.1 Pro or MiniMax M3.1 Flash if accuracy matters most and you can wait minutes per question. Skip Muse Spark Open for this workload, since it is one point behind and costs about three times as much per run. The &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;Artificial Analysis score of 31&lt;/a&gt;, against 29 for Inkling and 30 for Inkling-Small, points the same way: a real step up, not a leap.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: Is Inkling 2 Small really open source?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The weights are open under the Apache 2.0 licence, as &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;announced on November 12, 2026&lt;/a&gt;. Training data and code are not covered by that announcement, so "open weights" is the accurate term.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How big is Inkling 2 Small?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; It has &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;594B total parameters with 17B active per token&lt;/a&gt;, and handles image and text inputs. The context window reaches 1M tokens according to &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Can I run Inkling 2 Small on one machine?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Yes, with limits. &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;Plotly&lt;/a&gt; says the 4-bit build needs about 356GB, which fits a 512GB Mac Studio, and the MXFP8 build runs on a single server node. The &lt;a href="https://thinkingmachines.ai/model-card/inkling-2-small/" rel="noopener noreferrer"&gt;model card&lt;/a&gt; lists a minimum of 500GB aggregated VRAM for the main checkpoint.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How much does the API cost in rupees?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; About Rs 28 per million input tokens and Rs 94 per million output tokens, converting the &lt;a href="https://plotly.com/blog/inkling-2-small-plotly-data-analytics-bench" rel="noopener noreferrer"&gt;$0.30 and $1.00 prices&lt;/a&gt; at &lt;a href="https://www.divybharat.com/markets/usd-to-inr" rel="noopener noreferrer"&gt;Rs 94.05 per dollar&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is it good for chat assistants?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Not ideal. Thinking Machines &lt;a href="https://thinkingmachines.ai/news/inkling-2-small/" rel="noopener noreferrer"&gt;says it is less optimised for chat&lt;/a&gt;, with presentation issues and overcautious safety responses. It suits analysis and agent work better.&lt;/p&gt;

</description>
      <category>inkling2small</category>
      <category>openweightmodels</category>
      <category>thinkingmachines</category>
      <category>mixtureofexperts</category>
    </item>
    <item>
      <title>Open Weights Speech Model: SOV-Transcribe-1 vs Voxtral</title>
      <dc:creator>Shamuddin</dc:creator>
      <pubDate>Wed, 18 Nov 2026 04:43:14 +0000</pubDate>
      <link>https://dev.to/aitecharchive/open-weights-speech-model-sov-transcribe-1-vs-voxtral-33mg</link>
      <guid>https://dev.to/aitecharchive/open-weights-speech-model-sov-transcribe-1-vs-voxtral-33mg</guid>
      <description>&lt;p&gt;On 12 November 2026 Salesforce released &lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;SOV-Transcribe-1&lt;/a&gt;, and it is now the most accurate open weights speech model for recorded audio. Artificial Analysis &lt;a href="https://x.com/ArtificialAnlys/status/2120917242439409788" rel="noopener noreferrer"&gt;scored it at 2.3% on its AA-WER Index&lt;/a&gt;, ranking it 7th of 44 models including proprietary ones, ahead of Mistral's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe (both 2.8%). The verdict: pick SOV-Transcribe-1 if you can host an 80 GB-class GPU and need accurate English transcripts; pick a hosted API or a smaller model if you cannot.&lt;/p&gt;

&lt;h2&gt;
  
  
  TL;DR
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;SOV-Transcribe-1 is Salesforce's first open-weight speech-augmented language model, released under Apache 2.0 on 12 November 2026 (&lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;Hugging Face&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;It is the top open-weight model on the AA-WER Index at 2.3%, and 7th of 44 overall (&lt;a href="https://x.com/ArtificialAnlys/status/2120917242439409788" rel="noopener noreferrer"&gt;Artificial Analysis&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;It has 26.6B total parameters but only 4.6B active (&lt;a href="https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1" rel="noopener noreferrer"&gt;model card&lt;/a&gt;), so memory, not compute, is the constraint.&lt;/li&gt;
&lt;li&gt;The BF16 weights alone come to about 53 GB (derived below), which rules out 12 GB, 16 GB and 24 GB cards.&lt;/li&gt;
&lt;li&gt;It is non-streaming and covers seven languages, so it suits recorded meetings and calls rather than live captions.&lt;/li&gt;
&lt;li&gt;Last verified: 13 November 2026.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Which open weights speech model is most accurate right now?
&lt;/h2&gt;

&lt;p&gt;On the independent benchmark, SOV-Transcribe-1 is. Salesforce's own write-up reports a 2.28% word error rate, against 2.81% for NVIDIA Nemotron 3.5 Transcribe and 2.77% for Mistral Voxtral Small (&lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;Salesforce on Hugging Face&lt;/a&gt;). Artificial Analysis rounds the two rivals to 2.8% each.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;AA-WER (Salesforce figures)&lt;/th&gt;
&lt;th&gt;Active parameters&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;SOV-Transcribe-1&lt;/td&gt;
&lt;td&gt;2.28%&lt;/td&gt;
&lt;td&gt;4.6B&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Mistral Voxtral Small&lt;/td&gt;
&lt;td&gt;2.77%&lt;/td&gt;
&lt;td&gt;24B&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;NVIDIA Nemotron 3.5 Transcribe&lt;/td&gt;
&lt;td&gt;2.81%&lt;/td&gt;
&lt;td&gt;Same as SOV-Transcribe-1&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Source: &lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;Salesforce blog&lt;/a&gt;. The blog describes Voxtral Small as 4B against 24B active, while the model card rounds SOV-Transcribe-1 to 4.6B active.&lt;/p&gt;

&lt;p&gt;The gap over Nemotron is about half a percentage point. That is a real lead on a benchmark of this size, but it will not look dramatic on a clean podcast. It matters more on messy audio, which is where the other results are worth reading.&lt;/p&gt;

&lt;h2&gt;
  
  
  How does it do on hard audio such as earnings calls and call centres?
&lt;/h2&gt;

&lt;p&gt;Better than the other open models, with a caveat on absolute numbers. On HALAS, a set of real, human-annotated earnings-call recordings, it reaches 25.69% WER. Qwen3-ASR 1.7B scores 26.01%, Voxtral Small 27.35%, Nemotron 3.5 Transcribe 31.59% and IBM Granite Speech 5.0 2B 34.51% (&lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;Hugging Face blog&lt;/a&gt;). A quarter of words wrong is still a lot, so hard audio still needs human review.&lt;/p&gt;

&lt;p&gt;Other published per-set results, as reported by &lt;a href="https://cryptobriefing.com/salesforce-sov-transcribe-1-open-weights-speech-model" rel="noopener noreferrer"&gt;Crypto Briefing&lt;/a&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;4.1% on Earnings22 (earnings calls)&lt;/li&gt;
&lt;li&gt;1.5% on VoxPopuli (parliamentary speech)&lt;/li&gt;
&lt;li&gt;1.8% on AA-AgentTalk (voice-agent conversations), level with Nemotron&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;On AppTek's call-centre dialogues, a long-form agent-and-customer set spanning 14 English accents, the model averaged 8.21% WER and had the lowest error on 8 of the 14 accents, including Australian, British, Indian and Singaporean English (&lt;a href="https://benchlm.ai/models/sov-transcribe-1" rel="noopener noreferrer"&gt;BenchLM&lt;/a&gt;). If you serve a multi-accent English customer base, that result is more informative than the headline number.&lt;/p&gt;

&lt;h2&gt;
  
  
  Does it get product names and jargon right?
&lt;/h2&gt;

&lt;p&gt;Largely, yes. On VoiceCodeBench, where a recording only passes if every target entity is correct, it records a 76.3% task success rate and 93.4% entity recovery. That places it first among open-weight models and second among 23 systems overall (&lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;Hugging Face blog&lt;/a&gt;).&lt;/p&gt;

&lt;p&gt;Entity accuracy is the metric that tends to hurt in practice: a transcript with one wrong function name or drug name can be worse than one with several wrong filler words. The model also accepts natural-language instructions to steer transcript style and supports context biasing, so you can pass domain terms in advance (same source). Treat that as a feature to test on your own vocabulary rather than assume.&lt;/p&gt;

&lt;h2&gt;
  
  
  What hardware does an open weights speech model like this need?
&lt;/h2&gt;

&lt;p&gt;This is where the lead comes with a cost. The &lt;a href="https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1" rel="noopener noreferrer"&gt;model card&lt;/a&gt; lists 26.6B total parameters, including the audio encoder, with 4.6B active per token. The architecture pairs a Whisper large-v3 audio encoder and an audio projector with a Gemma 4 26B A4B mixture-of-experts language model. It ships in BF16, with the router in FP32.&lt;/p&gt;

&lt;p&gt;Our arithmetic, derived from those figures rather than stated by Salesforce: 26.6 billion parameters (&lt;a href="https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1" rel="noopener noreferrer"&gt;model card&lt;/a&gt;) at 2 bytes each is about 53 GB for weights alone, before the audio activations and cache. A mixture-of-experts design keeps compute low because few experts fire per token, but every expert must still sit in memory.&lt;/p&gt;

&lt;p&gt;Practical targets are:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;One 80 GB card (A100 or H100).&lt;/li&gt;
&lt;li&gt;Two 48 GB cards in tensor parallel, giving 96 GB.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;A 12 GB consumer card, a 24 GB card or a free Colab T4 with 16 GB cannot hold the BF16 build. As of 13 November 2026 the repository lists no official quantised or GGUF build. A community 4-bit conversion could change that, but none exists yet, so do not plan around one.&lt;/p&gt;

&lt;p&gt;Inference runs in vLLM (the repository includes a plugin for 0.32.0) and in Transformers, where native support is being upstreamed. The authors also publish SOV-Transcribe-1-DFlash, a draft model for speculative decoding that proposes 15 tokens at a time and speeds up single-request decoding by about 2 to 3 times, with accuracy within run-to-run variation (&lt;a href="https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1" rel="noopener noreferrer"&gt;model card&lt;/a&gt;).&lt;/p&gt;

&lt;h2&gt;
  
  
  What are the limits of SOV-Transcribe-1?
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Non-streaming only. It is built for recorded files such as meetings, calls and podcasts. If you need live captions, look elsewhere.&lt;/li&gt;
&lt;li&gt;Seven languages: English, French, German, Spanish, Italian, Portuguese and Dutch. The strongest evidence is for English.&lt;/li&gt;
&lt;li&gt;Audio is 16 kHz input, with up to 30 minutes tested per file (&lt;a href="https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1" rel="noopener noreferrer"&gt;model card&lt;/a&gt;). Longer recordings need chunking.&lt;/li&gt;
&lt;li&gt;Rankings are a snapshot as of 12 November 2026 and will move as other labs respond.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Should you self-host it or use a hosted service?
&lt;/h2&gt;

&lt;p&gt;Self-host if you have the GPU and a reason: data that cannot leave your network, steady high volume, or a need to tune the pipeline. The Apache 2.0 licence lets you do that commercially. The same logic applies to other open models we have covered, such as &lt;a href="https://dev.to/articles/longcat-2-5-open-weights-mit-self-host"&gt;LongCat 2.5 under an MIT licence&lt;/a&gt; and the &lt;a href="https://dev.to/articles/plamo-3-translate-open-weights-offline-53-languages"&gt;offline PLaMo 3 translation model&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you have a 12 GB to 24 GB GPU, or only occasional transcription, a hosted tool is the more sensible choice today. Our &lt;a href="https://dev.to/articles/best-ai-transcription-software-compared-2026"&gt;comparison of transcription software&lt;/a&gt; covers those options and their pricing trade-offs. For the speech-output side of open weights, see &lt;a href="https://dev.to/articles/zonos3-mini-open-source-tts-voice-cloning-browser"&gt;Zonos3 Mini for browser-based voice cloning&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Q: Is SOV-Transcribe-1 really open source?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The weights are published on Hugging Face under the Apache 2.0 licence, which permits commercial use. That makes it an open weights speech model; training data and full training code are not the same thing as the weights, so check the repository for what is included.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Can I run SOV-Transcribe-1 on a 24 GB GPU?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; Not in the shipped BF16 build. The weights alone come to about 53 GB by our arithmetic from the 26.6B parameter count, and no official quantised build is listed as of 13 November 2026.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: How much better is it than Voxtral Small and Nemotron 3.5 Transcribe?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; On the AA-WER Index it scores 2.28% against 2.77% for Voxtral Small and 2.81% for Nemotron 3.5 Transcribe, according to Salesforce's published figures. The margin is under a percentage point.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Does it support live, streaming transcription?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; No. It is designed for non-streaming transcription of recorded audio, so it is a poor fit for live captions or real-time voice agents.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Which languages does it cover?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; English, French, German, Spanish, Italian, Portuguese and Dutch, seven in total, according to the model card.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Q: Is the benchmark ranking independent?&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;A:&lt;/strong&gt; The AA-WER result comes from Artificial Analysis, an independent evaluator. Other results, such as HALAS and VoiceCodeBench, are reported by Salesforce and should be reproduced on your own audio before you commit.&lt;/p&gt;

&lt;h2&gt;
  
  
  Sources and corrections
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1" rel="noopener noreferrer"&gt;Salesforce announcement on Hugging Face&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1" rel="noopener noreferrer"&gt;SOV-Transcribe-1 model card&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://x.com/ArtificialAnlys/status/2120917242439409788" rel="noopener noreferrer"&gt;Artificial Analysis post on the AA-WER Index&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Last verified: 13 November 2026. Corrections log: none yet. This article was produced with AI assistance and reviewed against the primary sources above.&lt;/p&gt;

</description>
      <category>speechrecognition</category>
      <category>openweights</category>
      <category>selfhosting</category>
      <category>voiceagents</category>
    </item>
  </channel>
</rss>
