Gadgets Now
News
Laptops & Pc
Meta Opens Muse Glimmer, A 30B Agent Model For Laptops

Link CopyTwitterLinkedinWhatsapp

Meta Opens Muse Glimmer, A 30B Agent Model For Laptops

Meta released Muse Glimmer on 10 August 2026, a 30-billion-parameter open-weight agentic model published under an Apache 2.0 licence. Quantised to under 20GB, it runs local agents on a single 24GB consumer GPU or a Mac. Meta also said open weights for its frontier model, Muse Spark 1.2, follow in the coming weeks.

Sahil Mohan Gupta

Published: Aug 10, 2026, 04:59PM IST

Prefer us on

Meta Opens Muse Glimmer, A 30B Agent Model For Laptops

Meta Superintelligence Labs published Muse Glimmer on 10 August 2026, a 30-billion-parameter dense model with open weights under an Apache 2.0 licence, built to run always-on local agents on a single consumer graphics card. The weights are live on Hugging Face today. Alexandr Wang, Meta's chief AI officer, set the hardware bar in a public post, saying Muse Glimmer "can run on 24GB of VRAM without losing agentic reliability". That figure puts the model inside one mid-range desktop card, and Wang added that open weights for Muse Spark 1.2, Meta's current frontier model, follow soon. Four months after Meta shipped its first fully closed model, the company is handing weights back to developers, and doing it in the middle of a Washington argument about whether that should stay legal.

Muse Glimmer specifications: parameters, licence and hardware

About The Author

Sahil Mohan Gupta

At heart, I am a storyteller drawn to the watershed moments that bend the technology landscape. I braid narrative with data, humanise statistics, and trace the arc from first spark to world-changing impact. My reportage, features and reviews are witty, sardonic, visual and vivid, using anecdote to illuminate rather than eviscerate. As a technology journalist with over sixteen years of experience, I have travelled the world and the seven seas, covered every major tech conference worth its lanyard, chronicled the defining breakthroughs of the last decade and a half, and played a pivotal role in launching some of India’s most important technology publishing platforms across web, print and TV. In my current role as Editor of Gadgets Now Studios, I bring that experience, instinct and editorial firepower to the table, with the mandate of scaling the brand to towering heights. When I am off the clock, I am usually lost in music, from underground electronic and progressive rock to stone-cold blues. I am also an incurable F1 nut, a hangover from my previous life as an auto journalist, and always game for a jam session with friends, where I do my best to make my guitar gently weep.

Specification	Muse Glimmer
Parameters	30 billion, dense
Licence	Apache 2.0, permissive
Released by	Meta Superintelligence Labs
Release date	10 August 2026
Size at full precision	Over 55GB
Size after quantisation	Under 20GB, at roughly 4-bit
Memory envelope	24GB or 32GB
Input	Text and images, interleaved
Languages	Over 100
Available at	Hugging Face
Trained from	Distillation on Muse Spark outputs

Every figure above comes from Meta's own release. Reporting diverges on one point, the teacher model: Meta's post says Muse Glimmer trained on Muse Spark's outputs through logit distillation, while wire coverage of the launch describes the model as a distilled version of Muse Spark 1.2. Meta left the version number out of its technical post.

What Muse Glimmer does as a local agent

An agent that manages a calendar, drafts replies and sorts files needs deep access to personal context, which is the argument for keeping it on the machine that holds the files. Meta trained and evaluated Muse Glimmer across eight capabilities aimed at that job.

End-to-end task completion covers full-task benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench and SWE-Bench, which measure whether a model works inside a scaffold, writes and debugs code, and resolves multi-turn requests from start to finish. Tool use covers function calls with precise schemas held across long workflows. Multi-step reasoning covers plans that stay coherent over long horizons.

Failure recovery is the capability that separates a demo from a working agent. When a tool call fails or returns something odd, Muse Glimmer is trained to diagnose the error and retry. Anyone who has watched a local model call the same broken endpoint six times in a row understands why Meta put that on the list.

Multimodal input arrives through a dedicated perception encoder, so an agent can take a screenshot, a chart or a scanned document alongside the conversation. Scaffold compatibility covers OpenClaw and other orchestration patterns, which matters because most developers already have a harness and want a model that drops into it. Controllable effort lets the caller trade quality against speed. The multilingual training spans more than 100 languages.

How Meta fitted a 30B model onto a laptop

At full precision, 30 billion parameters need over 55GB of memory, which is more than any consumer graphics card offers. Meta compressed the weights to about 4-bit precision, taking the language model under 20GB. That leaves headroom inside a 24GB or 32GB envelope for the working memory, the perception encoder and the drafter model to run at the same time. Meta says quality holds through the compression on agentic tasks.

The second optimisation goes at generation speed. Language models produce text one token at a time, which drags during long reasoning chains and multi-step tool calls. Muse Glimmer ships with a lightweight drafter based on DFlash, a small companion network that proposes whole blocks of tokens at once. The main model checks those proposals in parallel, keeps the correct tokens and fixes the wrong ones. Output quality stays identical to standard generation. Meta includes quantised drafter versions so the drafter itself takes a smaller bite out of memory.

The audio equivalent is a pedal that adds headroom while leaving the tone alone. Compression buys space, the drafter buys speed, and the output reaching the user matches what the uncompressed model would have produced.

Device	Decode speed gain with DFlash
RTX 5090	3.1x
MacBook M5 Max	1.8x
MacBook M4 Max	1.5x

Those figures come from Meta's published chart, measured on a 17GB K-Quant build alongside the quantised drafter. The gap between 3.1x on a desktop card and 1.5x on last year's MacBook is a memory-bandwidth story, and it decides whether a local agent feels conversational or feels like waiting for a page to load in 2004.

Muse Glimmer against Gemma4, Qwen3.6 and the evaluation gap

Meta positions Muse Glimmer against Gemma4-31B and Qwen3.6-27B, and says it performs strongly for its size class across agentic, coding, multimodal, safety and reasoning categories.

The numbers behind that claim sit in an image. Meta published its evaluation as a table rendered as a picture, with the underlying detail pushed to a linked report instead of into the post itself. Developers who want to compare scores line by line have to go and fetch them. That is common practice across model launches and it works in the vendor's favour, because a claim of strong size-class performance travels further than a row of figures a rival can pick apart.

The honest position on day one: Meta's claim is Meta's, the third-party evaluations have yet to land, and the open weights mean anyone can check.

Three checks matter more than a leaderboard place. The first is whether the 4-bit build holds the quality Meta reports for the full-precision model, since almost every user runs the compressed version and the compression is where quality tends to leak. The second is failure recovery under real tool errors rather than benchmark ones, because a local agent meets broken APIs, rate limits and malformed responses all day. The third is long-horizon stability, the point at which a plan that started coherent drifts into loops. Vendors publish the first. The second and third stay in the community's hands. Wait for the community runs on τ-Bench and SWE-Bench before treating the size-class claim as settled.

Muse Spark 1.2 open weights: the bigger release Meta buried

The 30B model is the announcement. The sentence attached to it matters more. Wang said Meta will release an open-weight version of Muse Spark 1.2, the company's latest foundation model, in the coming weeks. Zuckerberg made the same commitment in his own post, writing that Meta will "release the weights for Muse Spark 1.2, our latest foundation model", and describing the company as a supporter of open source. Both posts left the date open.

	Muse Glimmer	Muse Spark 1.2
Weights	Open, Apache 2.0, available now	Open version promised in coming weeks
Size	30 billion parameters	Frontier-scale, undisclosed
Runs on	Single consumer GPU or Mac	Cloud infrastructure
Built for	Always-on local agents	General frontier capability
Access today	Download from Hugging Face	Paid API, launched 9 July 2026
Role in training	Student	Teacher

Meta has yet to say which version of Muse Spark 1.2 goes open, what licence it carries, or when it ships. Those three gaps decide whether the promise counts as a return to open weights at the frontier or as a smaller gesture attached to a model Meta has already superseded. The company made a similar-sounding commitment in the past and then shipped a closed model instead.

Why Meta reopened four months after going closed

Meta shipped Muse Spark on 8 April 2026, its first fully closed, proprietary model, built by Meta Superintelligence Labs under Wang after the company put $14.3 billion into Scale AI for a 49 per cent stake. The Llama series, which had passed 1.2 billion downloads by early 2026 and made Meta the loudest corporate voice for open weights, stopped being Meta's frontier offering. On 9 July, Muse Spark 1.1 opened a metered paid API, Meta's first pay-as-you-go model access service, priced below the incumbents.

For roughly four months the company that spent years arguing open weights spread power rather than concentrate it was selling access to a closed frontier model like every other lab. Today's release walks that back at the small end and promises to walk it back at the large end.

Three pressures explain the turn, and they compound. Chinese labs took the open-weights lead while Meta stayed closed, with Moonshot AI's Kimi K3 topping several benchmarks as an open-weight release. External developers building on open Meta models feed tools back into Meta's own products, which is a return Meta forfeits when the weights stay shut. And an open model that runs on the user's own hardware costs Meta zero in inference, which matters for a company spending at data-centre scale.

The competitive point is the sharpest of the three. Meta built its reputation on being the American lab that shipped weights. Losing that position to Beijing while also losing the API price war leaves the company short of both prizes.

The argument Meta has been making about who should hold AI

Zuckerberg has run the same line since his July 2025 letter on personal superintelligence, which set Meta's approach against rivals he described as directing superintelligence centrally at automating all valuable work. His position is that the useful version of this technology sits with the individual and answers to that person's goals.

Meta turned the argument into a campaign in late July 2026 under the line "The Future Is for Everyone", pitched against the doom messaging coming from other labs, and Zuckerberg extended it into a Wall Street Journal opinion column days later. The column works as the consumer-facing half of the open-weights lobbying push the company is already running with Nvidia, Microsoft and Palantir.

A model that runs on the user's own laptop is the cleanest expression of that position. The weights sit on your disk, the inference happens on your silicon, the files stay where they are, and the traffic stops at your machine. Whatever Meta gains from the release, it gains through developer goodwill and product feedback instead of through a metered API.

The qualification belongs in the same paragraph. Zuckerberg has signalled that Meta will keep some superintelligence models closed, and personal is a different word from open. Muse Spark spent four months as the proof of that, and Muse Spark 1.2 has yet to ship its open weights. The argument and the shipping record have been pointing in opposite directions all year, and today moves one of them.

The Washington fight this release is aimed at

Zuckerberg used the launch to call for lower US barriers on open-source models so American labs can compete with Chinese ones. That line lands inside a policy campaign already running.

On 24 July 2026, twenty-five organisations published an open letter titled "Open Weights and American AI Leadership", urging Washington to avoid premature restrictions on downloadable models. Signatories spanned the stack: Nvidia, Microsoft, Meta, Dell, IBM, Palantir, Mistral, Hugging Face, Mozilla, the Linux Foundation, Andreessen Horowitz and Y Combinator. OpenAI and Anthropic were absent from the original list. OpenAI added its name later as the roster passed thirty-five. Anthropic stayed off it.

The split follows the money. Companies selling chips, cloud capacity and open-weight models want the widest possible market. Companies selling closed frontier access through an API benefit when the alternatives face restrictions. Both camps describe their position as the safe one.

The trigger was distillation. US officials accused Moonshot AI of running large-scale distillation against American models to build Kimi K3, and the Treasury Secretary raised the prospect of sanctions on Chinese firms doing it. The coalition's answer is that distillation is a technique the whole field depends on, including Meta, which built Muse Glimmer by distilling its own frontier model, and that misappropriation deserves targeted enforcement instead of a blanket ban on open weights.

Washington has moved on the enforcement track in parallel. The US science and technology policy chief accused Moonshot AI of building an internal platform for large-scale distillation against American models, the Treasury Secretary raised sanctions as a response, and Nvidia, Microsoft, SpaceX and Palantir launched a separate security grouping in late July after a cyberattack on a leading American lab. The policy question has split into two: what to do about Chinese labs copying American model outputs, and whether American open weights should face restrictions as collateral. The coalition wants those kept apart. Some in Washington want them treated together.

Shipping a capable open model on the day you make that argument is a stronger move than making the argument alone. Muse Glimmer is the exhibit.

The case against open weights, taken seriously

Security researchers and several safety-focused labs make an argument that survives contact with the evidence: once weights are public, the safety training can be stripped. Fine-tuning a downloaded model to remove refusal behaviour costs a fraction of what training it cost, and the modified copy circulates beyond any vendor's reach. A closed API can revoke a key, log abuse and patch a jailbreak overnight. An open checkpoint on a hard drive answers to whoever holds it. Meta's own assessment framework has authority only up to the moment the file leaves Meta's servers.

That is why the safety-first labs stayed off the July letter. Anthropic declined to sign at all, and OpenAI joined late, once the roster had grown past thirty-five names. Their position is that staged release and centralised testing catch failure modes that a distributed release leaves standing.

The counter-argument holds partway. Wide release does harden defences faster in some categories, because thousands of researchers probing public weights find flaws sooner than any internal red team. It works less well for the categories where a single successful misuse is the harm, since the defensive benefit arrives after the fact.

The commercial symmetry is what the open-weights coalition leaves alone. Chip makers, cloud providers and open-model developers benefit when downloadable models stay legal, exactly as closed-frontier labs benefit when they face restrictions. Both camps describe a business interest as a safety principle, and both are partly right. Meta distilled Muse Glimmer from its own frontier model while arguing that distillation deserves protection from a blanket ban. That argument is self-serving and correct at the same time.

Running Muse Glimmer in India: the hardware floor priced in rupees

The download is free. The machine is where the cost sits, and in India that cost is steep.

Hardware that clears the bar	India price, indicative
MacBook Pro, M5 Max, 36GB unified memory	About Rs 5 lakh
Desktop RTX 5090, 32GB	Roughly Rs 4 lakh to Rs 5 lakh, with wide seller variance
RTX 5090 laptop, 24GB	Rs 6 lakh and above

Seller quotes for the desktop card spread far enough that a single figure would mislead, so treat the band as indicative and check current listings before budgeting. The direction holds regardless: a machine that runs Muse Glimmer at the speeds Meta measured costs more than most Indian small businesses spend on a year of software.

That gap changes who the release serves here. A funded startup in Bengaluru or Gurugram can buy the hardware and gain something real: an agent that handles customer records, financial documents or health data with the weights and the data staying inside the office, which sidesteps both API bills in dollars and questions about where personal data travels. An individual developer in a tier-two city gets a model they can read, fine-tune and study, running slowly on whatever they own.

India also has its own answer in the same size class. Sarvam AI open-sourced Sarvam 30B and Sarvam 105B in February 2026 under Apache 2.0, trained on IndiaAI Mission compute and tuned across 22 Indian languages. Sarvam 30B is a mixture-of-experts design with roughly 2.4 billion active parameters per token, which makes it lighter to serve than a 30B dense model like Muse Glimmer, and it holds a decisive edge on Indic-language evaluation. The IndiaAI Mission behind it runs on a Rs 10,372 crore allocation, with subsidised GPU access for startups.

Data residency sharpens the case for local inference here. India's data protection regime pushes firms handling personal data towards keeping processing under their own control, and sectors like lending, insurance, diagnostics and legal services carry client records that leadership teams prefer to keep off a foreign API. A model running on an office workstation answers that concern in a way a hosted endpoint struggles to, whatever the contract says.

For most Indian teams the practical route on day one avoids the hardware bill altogether. Together AI, Fireworks AI and OpenRouter offer hosted access to Muse Glimmer, which means a startup can test whether a 30B agent handles its workflow before committing five lakh rupees to a machine. Smaller quantisations run on cheaper cards at lower speed, which suits batch work and development, and works poorly for a live customer-facing agent or a coding assistant somebody waits on.

An Indian developer building a vernacular agent now picks between a Meta model tuned for tool use and failure recovery, and a domestic model tuned for the languages the customer speaks. Muse Glimmer's multilingual training spans over 100 languages, and Meta has yet to publish results for Hindi, Tamil, Bengali or Marathi. That measurement is the one Indian teams should run first.

Who should download Muse Glimmer today

Developers already building agents with a scaffold in place have the clearest case. Muse Glimmer works with OpenClaw and other orchestration patterns, integrations for llama.cpp, MLX and ExecuTorch arrive within days, and Ollama, LM Studio and Unsloth cover local running. Teams serving at scale get vLLM and SGLang, and Together AI, Fireworks AI and OpenRouter offer hosted access for anyone who wants the model minus the hardware bill. PyTorch's TorchTitan handles custom training. Meta is working with AMD, Arm, Dell, Intel and Nvidia on device-level tuning, and has published setup guidance for custom scaffolds through its developer documentation.

Anyone considering a hardware purchase to run this locally should wait. Independent benchmarks land within weeks, Muse Spark 1.2 weights land in the same window, and buying a Rs 5 lakh machine before either arrives is a decision made on a vendor's own numbers.

Everyone else gets to watch the more interesting question resolve. Meta has spent four months proving it can sell closed models and one day proving it can still ship open ones. Which of those two Meta shows up over the next twelve months settles what today meant, and the answer arrives with Muse Spark 1.2 instead of with anything written about Muse Glimmer this week.

FAQs

What is Muse Glimmer?

Muse Glimmer is a 30-billion-parameter dense AI model released by Meta Superintelligence Labs on 10 August 2026 with open weights under an Apache 2.0 licence. It is built for always-on local agent workflows including tool calling, local coding and LLM-as-a-judge evaluation, and runs on a Mac or PC with a single consumer GPU.

What hardware do I need to run Muse Glimmer?

A machine with 24GB or 32GB of available memory. At full precision the model needs over 55GB, so Meta quantised the weights to roughly 4-bit, taking the language model under 20GB and leaving headroom for the working memory, the perception encoder and the drafter. Meta measured performance on MacBook M4 Max and M5 Max machines and on an RTX 5090.

How fast is Muse Glimmer on consumer hardware?

Meta's published chart shows the DFlash speculative decoding drafter raising decode speed by 3.1x on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max, measured on a 17GB K-Quant build with the quantised drafter.

Is Muse Glimmer free to use commercially?

The weights are published under Apache 2.0, a permissive licence that allows commercial use, modification and redistribution. Meta has released the weights on Hugging Face along with developer documentation.

Will Meta open source Muse Spark 1.2?

Meta says yes. Alexandr Wang and Mark Zuckerberg both said on 10 August 2026 that an open-weight version of Muse Spark 1.2, Meta's latest foundation model, will follow in the coming weeks. Meta has yet to publish the licence, the exact configuration or a date.

How does Muse Glimmer compare with Indian open models like Sarvam 30B?

Both carry Apache 2.0 licences and sit in the same size class. Sarvam 30B is a mixture-of-experts model with roughly 2.4 billion active parameters per token, trained on IndiaAI Mission compute and tuned across 22 Indian languages, which makes it lighter to serve and stronger on Indic benchmarks. Muse Glimmer is a dense model tuned for agentic tool use, multimodal input and failure recovery, with over 100 languages in training and no published Indic-specific results.

end of article