NVIDIA Open-Sources Nemotron 3 Ultra 550B — Top US Open-Weights Model with 1M-Token Context and Commercial License

One-line summary: NVIDIA open-sourced Nemotron 3 Ultra, a 550B-parameter (55B active) hybrid Mamba-Transformer MoE model built for long-horizon agentic tasks — the strongest US open-weights release to date by benchmark intelligence score.

Key facts

• Architecture: 550B total / 55B active parameters; interleaved Mamba-2, MoE, and selective Attention layers (LatentMoE design)
• Context: 1 million tokens natively
• Throughput: 300+ tokens/second
• License: Linux Foundation OpenMDW-1.1 (commercially permissive)
• Score: 48 on Artificial Analysis Intelligence Index — #1 among US open-weights, behind China's Kimi K2.6 at 54
• Ships with 4 checkpoints (NVFP4, BF16 instruct, BF16 base, GenRM) plus training data and recipes
• Available June 4 on Hugging Face, OpenRouter, and NVIDIA NIM

Why it matters

With Meta's Llama release cadence slowing, NVIDIA is staking a direct claim in the open-weights ecosystem — not just as chip supplier but as model provider. Nemotron 3 Ultra beats other US open models by a wide margin on agent workloads, but the gap with China's open frontier (Kimi K2.6, GLM-5) signals that US open-source still trails internationally, keeping the competitive pressure high.

Read more

• NVIDIA AI Releases Nemotron 3 Ultra: Open 550B MoE Hybrid — MarkTechPost
• Nemotron 3 Ultra: high-speed, leading US open weights intelligence — Artificial Analysis