NVIDIA Nemotron 3 Ultra
Language modelOpen weightsCard summarizedFrontier reasoning, complex agentic workflows, long-context analysis, tool use, multilingual reasoning, high-stakes RAG
GA (v1.0)Source: Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning (technical report) + Hugging Face model card (65 pp)
Use it via API
Where to call it
- Self-host (weights)
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 - NVIDIA API catalog (build.nvidia.com)
nvidia/nemotron-3-ultra-550b-a55b
Good for
- Complex multi-step agents and long-running autonomous tasks
- Long-context reasoning over very large documents and codebases
- Code, math and science reasoning
Overview
Nemotron 3 Ultra
Frontier reasoning, complex agentic workflows, long-context analysis, tool use, multilingual reasoning, high-stakes RAG
- Cost
- $0.50 host
- Context
- 1M
- Cheapest host
- $0.50 / $2.20
- Max output
- —
- Knows up to
- September 2025
- Released
- 4 Jun 2026
The largest model in NVIDIA's Nemotron 3 family: a 550B-total, 55B-active hybrid Mamba-2/attention LatentMoE model with multi-token prediction. It was pre-trained in NVFP4 on 20T tokens and supports up to 1M tokens of context. NVIDIA releases the weights, most training data and the recipes under OpenMDW 1.1. NVIDIA reports accuracy on par with GLM-5.1, Kimi-K2.6 and Qwen-3.5 on agentic and reasoning benchmarks, with up to 5.9× higher inference throughput at 8K in / 64K out. p.1; HF model card
nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4- Input / output per 1M
- $0.50 / $2.20 via DeepInfra
- Context window
- 1M
- Max output
- —
- Knows up to
- September 2025
- Reasoning
- Configurable: on (default), medium effort, off, or a token budget
- Weights
- Open · OpenMDW License Agreement 1.1
About this card. A model card is the maker's own account of a model: what it's for, how it was trained and tested, and where it falls short. This page condenses NVIDIA's Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning (technical report) + Hugging Face model card. Every claim carries its source page.