Nemotron 3

NVIDIA's open hybrid Mamba-Transformer MoE pair: Nano 30B-A3B and Super 120B-A12B, 1M context, permissive license, open data.

In short. Nemotron 3 is NVIDIA's first family of open models built for AI agents. The small one, Nano, runs on a single good GPU. The larger one, Super, needs a server with several GPUs. Both read very long inputs, can switch their reasoning on or off, and can be used commercially. NVIDIA has since released the newer Nemotron 3.5.

Summary

NVIDIA announced Nemotron 3 on 15 December 2025 with Nano 30B-A3B, and added Super 120B-A12B on 11 March 2026. Both are hybrid models that interleave Mamba-2 state-space layers, a few attention layers and mixture-of-experts layers. Nano has 31.6 billion parameters and about 3.6 billion active per token (3.2 billion without embeddings). It was pretrained on 25 trillion tokens. Super has 120 billion parameters, 12 billion active, and uses LatentMoE, which routes tokens through a smaller latent space, plus multi-token-prediction layers for faster generation. It was pretrained on more than 25 trillion tokens in NVFP4. Both support up to 1 million tokens of context and a switchable reasoning mode with a thinking budget.

NVIDIA's cards compare Nano with Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B, and Super with Qwen3.5-122B-A10B and GPT-OSS-120B. NVIDIA's own results show Nano ahead of both on LiveCodeBench and up to 3.3 times the throughput of similar open models on one H200. Super leads on some math and long-context rows but trails Qwen3.5-122B-A10B on SWE-Bench (OpenHands) 60.5 against 66.4 and on TauBench V2 (61.2 against 74.5). These are NVIDIA's own tests.

The release is unusually open: NVIDIA published weights, the training recipe and most of the training data. Nano needs about 63 GB in BF16 or 19 GB in NVFP4, and Super about 247 GB in BF16 or 80 GB in NVFP4; the card lists eight H100 80 GB GPUs as the minimum for Super. The license is the NVIDIA Nemotron Open Model License, which allows commercial use. NVIDIA's Nemotron 3.5 generation, with Lightning 30B-A3B in August 2026 and Super VL 120B-A12B in October, is the successor to both sizes.

Good for

  • Single-GPU agent work with Nano, especially in NVFP4
  • High-volume agent workloads such as ticket automation with Super
  • Long-context RAG and document work up to 1 million tokens
  • Teams that want open weights together with open data and recipes

Watch out for

  • Nemotron 3.5 is newer and succeeds both sizes, so compare it for new projects
  • Super needs a multi-GPU node, with eight H100 80 GB GPUs as the stated minimum
  • The context window is advertised up to 1M tokens, but the default configuration of Nano uses 262,144
  • Benchmark claims come from NVIDIA's own evaluations

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Nano 30B-A3B (NVFP4) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Nano 30B-A3B (BF16) BF16 31.6B 3.6B active 73 GB 36 GB 22 GB Gaming GPU
Nano 30B-A3B (NVFP4) NVFP4 31.6B 3.6B active 73 GB 36 GB 21 GB Gaming GPU
Super 120B-A12B (BF16) BF16 120B 12B active 276 GB 138 GB 83 GB Workstation
Super 120B-A12B (NVFP4) NVFP4 120B 12B active 276 GB 138 GB 88 GB Workstation

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

A permissive, Apache-style license from NVIDIA. Commercial use, derivative works and redistribution are allowed, NVIDIA claims no ownership of outputs, and the main duties are to pass the license on and keep notices.

More about NVIDIA Nemotron

Frequently asked questions

What is Nemotron 3?
Nemotron 3 is NVIDIA's first family of open models built for AI agents. The small one, Nano, runs on a single good GPU. The larger one, Super, needs a server with several GPUs. Both read very long inputs, can switch their reasoning on or off, and can be used commercially. NVIDIA has since released the newer Nemotron 3.5.
Can I use Nemotron 3 commercially?
Yes. Nemotron 3 is released under NVIDIA Nemotron, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Nemotron 3 open source?
Nemotron 3 is open-weight, not open source in the strict sense. You can download the model files, but the NVIDIA Nemotron adds conditions that the Open Source Initiative definition does not allow.
Can I run Nemotron 3 on my own computer?
Yes, on a good graphics card. The smallest variant, Nano 30B-A3B (NVFP4), needs about 21 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was Nemotron 3 released?
Nemotron 3 was released by NVIDIA on March 11, 2026.
How long a context window does Nemotron 3 have?
Nemotron 3 supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Nemotron 3.5 Lightning NVIDIA 31.6B 3B active 1M OpenMDW-1.1 (Permissive) 22 GB Gaming GPU
Nemotron 3.5 NVIDIA 8B – 120B 12B active 1M OpenMDW-1.1 (Permissive) 5.52 GB Laptop or phone
Nemotron 3 Ultra NVIDIA 550B 55B active 1M OpenMDW-1.1 (Permissive) 379 GB Big-memory desktop
Nemotron 3 Nano Omni NVIDIA 31B 3B active 256K NVIDIA Open Model (Permissive) 32 GB Gaming GPU