본문으로 건너뛰기
All news

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

뉴스레터 구독

무료 뉴스레터

매주 핵심 AI 소식, 한 번에 받기

쏟아지는 AI·LLM 뉴스 중 꼭 알아야 할 것만 골라 메일로 보내드려요. 뉴스레터 발송이 시작되면 구독자분들께 가장 먼저 보내드립니다.