NVIDIA's Nemotron 3 Ultra is now the strongest open US model — 550B parameters, built for agentic coding workflows

NVIDIA announced Nemotron 3 Ultra at Computex 2026 on June 1, as part of Jensen Huang's keynote in Taipei. It is a 550-billion-parameter mixture-of-experts model with approximately 55 billion active parameters per token — meaning roughly 90% of parameters are inactive at inference time, which is the architectural reason for its speed. Weights are available from June 4, 2026 on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as NVIDIA NIM microservices, with additional availability through cloud partners including DeepInfra.

Where it ranks. Artificial Analysis's independent evaluation — conducted in partnership with NVIDIA — places Nemotron 3 Ultra at 48 on its Intelligence Index, making it the highest-scoring US open-weights model on that leaderboard. For context: Gemma 4 31B scores 39, Nemotron 3 Super scores 36, and gpt-oss-120b scores 33. The honest caveat: it does not top the global open-model chart. China's Kimi K2.6 scores 54 on the same index, and the closed Claude Opus 4.8 scores 61. Nemotron 3 Ultra is the best open US model — not the best open model overall.

Speed. Pre-release testing on DeepInfra shows Nemotron 3 Ultra delivering over 300 tokens per second — materially faster than comparable Chinese open models like DeepSeek and Moonshot Kimi at 50–100 tokens per second. NVIDIA also claims 5× faster inference and up to 30% lower cost for complex agentic tasks versus "open frontier models in its class," though the specific comparison models and methodology behind those figures are not published.

What NVIDIA optimised it for. Nemotron 3 Ultra was post-trained specifically for agentic harnesses rather than general-purpose chat. NVIDIA lists compatibility with Hermes Agent, LangChain Deep Agents, OpenClaw, OpenHands, and OpenCode — the orchestration layers that developers use to run autonomous multi-step coding and research workflows. The practical implication: if you already run agents on one of those harnesses, Nemotron 3 Ultra is a drop-in model swap. Weights ship in BF16 and NVFP4 quantization formats. Pricing is not yet disclosed.

The open-model context. Nemotron 3 Ultra arrives alongside NVIDIA's other Computex launch, Cosmos 3 (announced May 31), which targets physical-AI and video generation. Nemotron 3 Ultra is the complementary language-and-reasoning bet — NVIDIA shipping open weights at both ends of the capability stack. For developers evaluating open models for agentic pipelines, it raises the ceiling for what is achievable without paying closed-model API rates. Whether the intelligence gap versus Kimi K2.6 matters depends on the task: at 300+ tokens per second and open weights, the economics of running high-volume agent loops favour Nemotron 3 Ultra even if raw intelligence benchmarks are not first globally.

Who it is for. Nemotron 3 Ultra is a developer and infrastructure tool, not a consumer product. It competes with other open models in coding-agent pipelines — the same tier as Claude Code (which uses closed Anthropic models) and Cursor (which uses a mix of models). The direct comparison is not product-to-product — it is model-to-model for developers who self-host or route through OpenRouter. If your workflow is built on an IDE like Cursor or Claude Code, nothing changes today. If you build agentic pipelines on OpenRouter or HuggingFace and want a fast, open US alternative to closed frontier models, Nemotron 3 Ultra is now the strongest candidate. See our AI coding tools roundup for the broader category landscape.