NVIDIA Nemotron
NVIDIA's open Nemotron 3 models (Nano to 550B Ultra), hybrid Mamba-Transformer MoE with 1M context, released with weights, data and training recipes
Overview
NVIDIA Nemotron is a family of open models with open weights, training data and recipes, aimed at building AI agents that reason. The current generation, Nemotron 3, uses a hybrid design that mixes Mamba-2 layers, Transformer attention and mixture-of-experts, with up to a 1M-token context. Nemotron 3 Super (March 2026) has 120B total and 12B active parameters and was pretrained natively in NVFP4; Nemotron 3 Ultra (June 2026) has 550B total and 55B active, and Artificial Analysis called it the best US open-weights model at launch. NVIDIA also lists Nano, Nano Omni (30B, 3B active, video, audio, image and text input), Lightning and Nemotron 3.5 Super VL tiers, plus Nemotron models for retrieval, speech and safety. Licenses differ by checkpoint: Super uses the NVIDIA Nemotron Open Model License ("ready for commercial use"), while the Ultra model card uses the OpenMDW License v1.1. NVIDIA published the training recipes on GitHub and released part of the training data (about 10 of 37 datasets for Super). Also from NVIDIA: NVIDIA NeMo Data Designer, the open-source synthetic data framework used to build Nemotron training data, Isaac GR00T, foundation models for robots, and OpenShell, a sandboxed runtime for AI agents.
The Verdict
Who Should Use Nemotron?
Best For
- Agent builders who need very long context (1M tokens) in an open model
- Teams on NVIDIA GPUs who want models tuned for that hardware (NVFP4)
- Researchers who want training recipes and open data, not just weights
- Enterprises that want a US-made open model
Not Ideal For
- Users who want few refusals (some find it "overly locked down")
- Top coding results (users rank Qwen3.5 ahead in some tests)
- Lowest hosted price per token (Artificial Analysis finds gpt-oss-120b cheaper and faster)
What's Great
- Open weights, recipes and part of the training data
- 1M-token context; NVIDIA reports 91.75 on RULER at 1M vs 22.30 for gpt-oss-120b
- Ultra scored 47.7 on the Artificial Analysis Intelligence Index at launch
- Greptile found Super effective at code review with few tool calls
Watch Out For
- Different licenses per checkpoint (NVIDIA Open Model License vs OpenMDW)
- Only some training data is public, so recipe results differ from the paper
- Users report frequent refusals and misuse flagging
- Super scores 60.47% on SWE-Bench Verified (NVIDIA figure), behind some rivals
Pricing
View all features & details
Nemotron 3 Models
- Ultra: 550B total, 55B active, up to 1M context (Jun 2026)
- Super: 120B total, 12B active, native NVFP4, 1M context (Mar 2026)
- Nemotron 3.5 Super VL: 120B, 12B active, vision-language
- Nano Omni and 3.5 Lightning: 30B, 3B active
- Bolt and Nano tiers (sizes not stated on the overview page)
Other Nemotron Families
- Retrieval: extraction, embedding, reranking
- Speech: ASR, diarization, TTS, translation (3.5 Transcribe)
- Safety (NemoGuard): jailbreak, moderation, PII, topic control
What Is Open
- Weights on Hugging Face and as NIM
- Training recipes and cookbooks on GitHub (Apache-2.0 repo)
- About 10 of 37 pretraining and post-training datasets for Super
Real-World Usage
How It Compares
| Feature | Nemotron 3 Ultra | Nemotron 3 Super | Gemma 4 31B | gpt-oss-120b |
|---|---|---|---|---|
| AA Intelligence Index | 47.7 | 36.0 | 39.2 | 33.3 |
| Parameters | 550B (55B active) | 120B (12B active) | 31B dense | 120B |
| Context | 1M | 1M | 256K | Not stated |
| Open training data | Partial | Partial | Not stated | Not stated |
| License | OpenMDW 1.1 | NVIDIA Open Model License | Apache 2.0 | Not stated |
Sources: NVIDIA Nemotron · Nemotron 3 Super (Mar 2026) · Ultra model card · build.nvidia.com · Artificial Analysis · GitHub
Is this your tool? Add a badge