Official NVIDIA Blog & Developer Blog RSS (AI-filtered) + GitHub + Hugging Face.
63 updates · page 1 of 5
NVIDIA describes Meta's Muse Spark Open, a coding-focused 816B mixture-of-experts model with 42B active parameters and a 1M token context window, trained on NVIDIA GB300 NVL72.
Why it matters: Teams running agentic coding on NVIDIA GB300 NVL72 can use Muse Spark Open's long context.
NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter model for specialized always-on agents, plus Bolt 74B-A9B 24GB and Bolt E4B 4GB edge models.
Why it matters: Developers can pick a Nemotron 3.5 model for always-on agents or a smaller Bolt model for local workloads.
NVIDIA Nemotron 3.5 Super VL is an open model for always-on agents that read documents, check sources, follow policy, and call tools.
Why it matters: Enterprise teams can run longer multimodal agent workflows on an open Nemotron 3.5 model.
NVIDIA is expanding advanced technology access for Europe's scientific community through open learning resources, collaborations, and compute.
Why it matters: Smaller labs and early-career scientists in Europe can get more of a path to accelerated computing.
NVIDIA NeMo Gym is used to improve agent reliability and tool-calling efficiency for chip design, where plausible but incorrect artifacts can block production use.
Why it matters: Chip-design teams can use NeMo Gym to catch incorrect agent artifacts before they propagate through multistage workflows.
NVIDIA announced new investments in the UK startup ecosystem, including a five-year, $250 million compute commitment for the UK's agentic safety and security research.
Why it matters: UK startups and researchers can look to that compute commitment for agentic safety and security work.
Cities and partners across EMEA are using NVIDIA accelerated computing to build physical AI agents and digital twins that help decide road repairs and energy use, and can anticipate maintenance needs.
Why it matters: City teams can use these NVIDIA agents and digital twins to anticipate maintenance and guide day-to-day urban decisions.
NVIDIA explains how to rewrite Apache Spark UDFs for GPU acceleration with AI agents as agentic workloads need more throughput on large datasets.
Why it matters: Data teams can follow the guide to move Spark UDFs onto GPUs with help from AI agents.
NVIDIA announced new investments in France's startups as companies there deploy agentic computing across space, financial services, the public sector, and scientific research.
Why it matters: Founders and labs in France can see NVIDIA's new investment behind that agentic computing work.
ReplaySSM rethinks state storage so hybrid models that mix state-space layers with a few attention layers can decode faster. NVIDIA Nemotron 3, Qwen3.5 through Qwen3.8, Kimi-K3, and GLM-5.3-Flash are built this way.
Why it matters: Serving stacks for those hybrid models can decode faster by changing how SSM state is stored.
At the OCP Global Summit, NVIDIA announced partnerships and contributions that open the Vera Rubin architecture as a common foundation for scaling AI factories.
Why it matters: Infrastructure teams can build on a shared Vera Rubin foundation as AI factories grow from megawatts to gigawatts.
TensorRT-LLM v1.4.0 notes known issues: TP2 streaming or RPC executor workloads may hang when MPI worker pools are reused (observed on H100), DeepSeek MLA models on SM120/SM121 may crash with BF16 BMM CUDA graphs, and GPT-OSS on SM120 may fail with attention data parallelism.
Why it matters: People upgrading to TensorRT-LLM v1.4.0 can check those H100, SM120, and SM121 failure cases before they hit them.
NVIDIA describes developers combining frontier AI models with Omniverse libraries so agents help assemble assets, connect physics and rendering, and check that a simulation behaves as intended.
Why it matters: Simulation developers can direct agents in natural language to explore scenarios, investigate failures, and improve designs.
NVIDIA outlines five steps to make CAD assets SimReady for robotics, including materials, collision geometry, joints, and other physics properties in OpenUSD before testing behavior in Omniverse.
Why it matters: Robotics developers can prepare simulation-ready assets instead of only converting CAD geometry.
NVIDIA's KGMON team placed second in the KDD Cup 2026 Data Agents competition with a smaller, clearer harness that is easier to verify, built to answer natural-language questions.
Why it matters: Teams building data agents can study a competition system that won by shrinking the harness rather than adding to it.