NVIDIA AI

What are you building with open models? Show us what you’re working on and you could be headed to #NVIDIAGTC Berlin. 🎫 Your Golden Ticket includes: → Free conference pass → VIP seating for Jensen’s keynote → Exclusive NVIDIA merch → Access to special events Submissions are open Aug 18 – Sep 10, 2026. Details and how to enter: https://nvda.ws/4ieFwCm

What goes into building an open model for long-running agents? In this Ask the Experts session, you'll get direct access to the NVIDIA AI researchers building and training the latest Nemotron open models. Join us to hear what's driving their design decisions: how these models are trained to handle specialized, high-volume agentic tasks, why Nemotron open models are built to run locally — on hardware like DGX Spark — without sacrificing accuracy or speed, and what it means to build an open model that fits into a system where different tasks call for different models. We'll cover how the research team approaches training open models purpose-built for agentic work: fast, efficient, and designed to be customized for your domain. The conversation will dig into the architecture choices, training techniques, and design principles behind models built to run anywhere — from DGX Spark to the data center — and to slot cleanly alongside other models in a multi-model agent system. What you'll learn: How NVIDIA AI researchers train open models for long-running, autonomous agents What makes the Nemotron family customizable for domain- and task-specific workflows How Nemotron open models are optimized to run locally, from DGX Spark to your own infrastructure How Nemotron open models are built to work in systems where different steps call for different models Have questions about open model training, agentic design choices, or running Nemotron locally? Drop them live — our researchers will answer them directly.

We just released TensorRT Model Connect in Public Preview. You can take a supported Hugging Face model to end-to-end TensorRT inference in just two commands. No intermediate ONNX export, and the resulting bundle can run through native C++ APIs. We also built the entire project with OpenAI Developers Codex agents, with humans directing and reviewing the work. That includes model implementations, performance tuning, tests, integrations, and docs. It’s open source, so go try it out, dig into the implementations, or contribute support for a new model: https://lnkd.in/dAHCMX8S

⏳T-Minus 7 Days Until #WorldEntrepreneursDay 🛠️ Matan Grinberg started Factory from a simple observation: the world has endless problems solvable with software, but building good software is really challenging. When early LLMs made program synthesis real, his team saw a way to fix that, and built the first demo in 72 hours. Factory AI wants software to stop being scarce. When Droid handles the drudgery of the development lifecycle end to end, a small team can ship what once took hundreds of engineers, and more problems can be solved with good software. 💡 His advice to founders: "The only way you can win is by executing faster and being more obsessed."

Ready to run Qwen3.8 locally? 👀 Qwen3.8-27B brings powerful AI to an open model you can download, run locally and build with on your own terms. With Day 0 support for NVIDIA RTX and Jetson, bring that intelligence to the edge at up to 131 tokens per second on a single GeForce RTX 5090.

Building capable robot foundation models requires more than one approach. Vision-language-action (VLA) models and World Action Models (WAMs) - each offer different strengths for robot learning, planning, and control. In this livestream, NVIDIA researchers Danfei Xu US Moritz Benno Reuss CH Thomas Tian US will explore what World Action Models work, how they compare with modern Vision-Language-Action (VLA) models, and why many next-generation robot foundation models combine both approaches. You'll also see how NVIDIA Cosmos 3 treats action as a native modality, enabling robots to predict future observations while generating actions. We'll walk through how Cosmos 3 provides an open workflow for post-training, evaluation, and deployment using open models, datasets, recipes, and serving resources. Whether you're building manipulation policies, training robot foundation models, or exploring embodied AI, this livestream provides a practical understanding of where World Action Models fit into the future of robotics. What You'll Learn: - What World Action Models are and how they differ from and complement Vision-Language-Action models - The strengths, limitations, and tradeoffs of WAMs, VLAs, and emerging hybrid approaches - How Cosmos3-Nano-Policy-DROID imagines future observations while generating robot actions - How NVIDIA Cosmos 3 represents action as a native modality alongside video - How to evaluate, post-train, and adapt Cosmos 3 using open models, datasets, recipes, and serving resources Have questions about how to post-train and deploy NVIDIA Cosmos 3? Drop them live — the NVIDIA team will answer them in real time. Access more NVIDIA Cosmos developer resources and join our developer community: 📄 Read the Technical Blog → https://nvda.ws/4c6kK3R ⬇️ Download Cosmos on Hugging Face → https://lnkd.in/gE_uy_jT 📚 Explore Models & Datasets on GitHub → https://lnkd.in/gRY3QEvU 👥 Join the Cosmos Community → https://lnkd.in/dpCPQSmj

Massive open-weight model drop! Congratulations to Qwen on releasing the open weights for Qwen3.8-2.4T-A95B, a 2.4T parameter model with 95B active, designed for demanding reasoning and agentic workloads. Out of the box, the model achieves a throughput of 4K+ tokens/second per GPU and 350+ tokens/second per user on NVIDIA GB300 NVL72 in FP8 precision. Read the tech blog: https://nvda.ws/463kFdP

Running frontier-like open models locally means no cloud dependency, no per-token costs after hardware, and all data staying on-device. At SIGGRAPH, NVIDIA announced the NVIDIA Agent Toolkit stack on DGX Station — NemoClaw, Nemotron 3 Ultra, Omniverse libraries, and OpenShell secure runtime in a single deskside supercomputer. This session unpacks what that stack means for developers and walks through a working example: deploying Nemotron 3 Ultra on DGX Station. We'll get hands-on — launching Nemotron 3 Ultra (550B, NVFP4) on a GB300 DGX Station with vLLM, including MoE expert CPU offloading, speculative decoding, and tool-call routing to an OpenAI-compatible API. And cover the SIGGRAPH announcements and what the full Agent Toolkit stack enables. What you'll learn: What the NVIDIA Agent Toolkit stack on DGX Station includes and why it matters for local AI How to serve Nemotron 3 Ultra on DGX Station using vLLM with NVFP4 quantization and CPU offloading How to configure speculative decoding, prefix caching, and tool-call routing for agentic workloads How to verify and query a locally running Nemotron 3 Ultra endpoint Running open models like Nemotron locally on DGX Station or DGX Spark? Bring your setup questions — we'll answer them live.

In this session we will focus on how to bring VLM/VLA models to power real-world physical AI applications. We will focus on how to utilize SOTA of VLM (gemma 4) and or GR00T model for performing different pick and place tasks and orchestrate the outputs to control the robots using ROS 2 framework. You will learn how to bring vision-language models into real-world physical AI applications — from model selection to robot control. We'll cover: Choosing the right model for robotics — learn when to use a state-of-the-art VLM like Gemma 4 versus a specialized model like NVIDIA GR00T, and how runtime, throughput, and task requirements shape that decision. VLMs and VLAs in action — see how vision-language and vision-language-action models are applied to real manipulation tasks like pick and place, and what makes them viable for physical AI. Connecting model outputs to robot control — understand how to orchestrate model outputs through the ROS 2 framework to drive real robot behavior. Hands-on hardware demo — walk through a live example using the SO-101 or reBot Arm, putting everything together from model inference to physical actuation.

Introducing NVIDIA Nemotron 3.5 Lightning ⚡ An open 30B MoE model with 3B active parameters, built for always-on agents to complete high-volume, specialized tasks faster. Lightning delivers up to 4x the output speed of similar-sized models.  And because not every step in an agent workflow needs the same model, we’re also releasing NVIDIA NeMo Switchyard, a new open source library for model routing. Use frontier models for complex reasoning and planning, and Lightning for high-volume, specialized execution. As always, Nemotron 3.5 Lightning is open and customizable, including model weights, training data and post-training recipes.