NVIDIA Nemotron 3 Ultra (550B-A55B)

2026-06-04 · Publisher : NVIDIA · Domain : Artificial Intelligence

NVIDIA releases Nemotron 3 Ultra, a 550-billion-parameter MoE with 55 billion active parameters for long-running agents, with weights, data and recipes under the OpenMDW-1.1 license.

### Context Ultra was announced in December 2025 at about 500 billion parameters; the model as delivered reaches 550 billion. It targets reasoning and the orchestration of long-running agents. NVIDIA compares it with GLM 5.1, Kimi K2.6 and Qwen 3.5, which are open models. ### What the model brings - Multi-teacher on-policy distillation (MOPD) with more than 10 specialized teacher models, in addition to supervised fine-tuning and reinforcement learning. - A claimed throughput of up to 5 times that of comparable open models, and 5.9 times that of GLM-5.1, 4.8 times that of Kimi K2.6 and 1.6 times that of Qwen-3.5 (8k input, 64k output). - A task completion cost up to 30% lower on SWE-bench and Terminal-Bench 2.0 according to NVIDIA. - Additional pretraining data (212 billion tokens, including 173 billion tokens of GitHub code up to September 30, 2025) and, in post-training, 10 million SFT samples, 1 million RL tasks and 15 RL environments. - An OpenMDW-1.1 license from the Linux Foundation, presented as permissive. ### Architecture and specifications 550 billion parameters in total, 55 billion active, Mamba-2 Transformer with LatentMoE and MTP, NVFP4 precision (Hopper, Blackwell, Ampere). Context up to 1 million tokens. Pretraining on about 20 trillion tokens (cutoff September 2025), post-training through May 2026. Languages: English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, Chinese. License: OpenMDW-1.1 (architecture, parameters, documentation and software), different from the NVIDIA Nemotron Open Model License of Nano and Super. Openness: weights (base, BF16, NVFP4, GenRM), four data collections, NeMo recipes (LoRA SFT, full SFT via Megatron Bridge, GRPO, Dynamo) and a technical report; the full corpus is not published, only a significant part of it is. ### Cost and access | Item | Detail | |---|---| | Official API | build.nvidia.com, Perplexity, OpenRouter, more than 20 platforms including Amazon SageMaker, Google Cloud and Microsoft Azure; no pricing disclosed | | Minimum hardware | 8 x B200/GB200/GB300 or 16 x H100 (HF model card) | | Weights | Base, BF16 and NVFP4 on June 3, 2026 on Hugging Face | ### Benchmarks | Benchmark (measured by NVIDIA) | Ultra | GLM 5.1 | Kimi K2.6 | Qwen 3.5 | |---|---|---|---|---| | PinchBench | 91% | 84% | 91% | 89% | | EnterpriseOps-Gym | 33% | 40% | 29% | 30% | | Terminal-Bench 2.0 | 54% | 64% | 67% | 53% | | IFBench | 82% | 77% | 74% | 78% | | GDPVal-AA | 1,448 | 1,594 | 1,508 | 1,192 | | ProfBench | 56% | 46% | 56% | 53% | | RULER 1M | 95% | not tested (256K) | not tested (256K) | 90% | Source: NVIDIA technical blog post of June 4, 2026, NVIDIA measurements. HF model card, other scores: MMLU-Pro 86.8, SWE-Bench Verified 70.7, LiveCodeBench 89.0, IOI 2025 570.0. The comparison models are open models. ### Versus closed models NVIDIA's table contains no closed model, and no comparison with a closed model from June 2026 was found in the sources consulted. Against open models, Ultra leads on IFBench (82% versus 74 to 78%) and on long context (RULER 1M at 95%), but is clearly behind on Terminal-Bench 2.0 (54% versus 64% for GLM 5.1 and 67% for Kimi K2.6) and on GDPVal-AA. NVIDIA mostly emphasizes throughput and cost per task. Open weights under OpenMDW-1.1 allow self-hosting, provided you have 8 Blackwell GPUs or 16 H100s. Any claim of parity with closed models comes from NVIDIA and is not substantiated here. ### Key takeaways Nemotron 3 Ultra places NVIDIA among the large open models, with a permissive license and published data. It requires a large GPU cluster.

Tags : LLM, Open Source, MoE, NVIDIA, Nemotron 3, OpenMDW-1.1, Poids, données et recettes ouverts