Nemotron 3 Ultra
NVIDIA
Key facts · verified Oct 30, 2026
Nemotron 3 Ultra is a NVIDIA model released on 4 Jun 2026. Pricing: Open weights, self-hosted. Its context window is 1M tokens. It ranks #120 on LMArena's text leaderboard (Elo 1427, read 1 Nov 2026). Source: huggingface.co.
NVIDIA's Nemotron 3 Ultra, the flagship of its open-weight Nemotron 3 family: a Mamba2-Transformer hybrid mixture-of-experts model with 550B total and 55B active parameters, a 1M-token context window, released under the OpenMDW-1.1 license. NVIDIA's own model card states it is optimized for complex multi-step agents and high-accuracy reasoning over code, math and science, reporting 70.7% on SWE-Bench Verified and 89.0% on LiveCodeBench (v6).
Pricing & specs checkedSource: vendor documentation · checked Oct 30, 2026LMArena: 1427 Elo (rank 120) · source · retrieved Nov 1, 2026
Key Features
Open Weights MoE 55B Active / 550B Total 1M Context Agentic Long-Context Reasoning Self-Hostable
Full Specifications
Open weights, self-hosted | |
|---|---|
| Identity | |
| Developer | NVIDIA |
| Released | Jun 2026 |
| Status | GA |
| Licence | Open weights (OpenMDW-1.1) |
| Self-hostable | Yes |
| Cost | |
| Input price | Self-hosted |
| Output price | Self-hosted |
| Cached input | Not offered |
| Batch discount | Not offered |
| Free tier | Free rate-limited NIM API endpoints (NVIDIA API Catalog / build.nvidia.com) for prototyping only, not production |
| Capacity | |
| Context window | 1M tokens |
| Capability | |
| Vision in | No |
| Audio in | No |
| Function calling | Yes |
| Structured output | Yes |
| Extended reasoning | Yes |
| Web search | Not offered |
| Code execution | Not offered |
| Access | |
| API | Yes |
| Chat app | Yes |
| Cloud marketplaces | Azure AI Foundry |
| Fine-tuning | Yes (self-hosted via NVIDIA NeMo: LoRA, SFT, RL) |
| Measured quality | |
| LMArena Elo | 1427 (checked Nov 2026) |
| LMArena rank | Rank 120 |
| GPQA Diamond | 85 |