News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnNemotron 3 Nano

Nemotron 3 Nano

NVIDIA's 30B hybrid Mamba-Transformer MoE model for reasoning and agents

Nemotron 3 Nano is a 30-billion-parameter language model trained from scratch by NVIDIA and released in December 2025. It mixes Mamba-2 layers, attention and a mixture-of-experts design so only 3.5B parameters are active per token, and it handles both reasoning and non-reasoning tasks for agent and chatbot work.

Share on WhatsApp

  • 30B total parameters but only 3.5B active per token
  • 52 layers mixing Mamba-2, MoE and attention
  • Trained on 25 trillion tokens, released 15 December 2025
  • Reasoning can be switched on or off in the chat template
  • Weights released under the NVIDIA Nemotron Open Model License

What it is

Nemotron 3 Nano is a large language model built by NVIDIA and put out on Hugging Face on 15 December 2025. It holds 30 billion parameters in total, but only about 3.5 billion of them are active for any one token, which keeps it efficient to run. NVIDIA describes it as a single model for both reasoning and ordinary answering: it can write out a thinking trace before giving a final reply, or skip that step when the user prefers. It is aimed at developers building agents, chatbots, retrieval systems and instruction-following tools, mainly in English and code, and it is cleared for commercial use.

How it came about

NVIDIA says the model was trained from scratch between September and December 2025. Pre-training data stops at 25 June 2025 and post-training data at 28 November 2025. It belongs to the Nemotron line of open models, following the Nemotron 2 family, and it appeared on Hugging Face on 15 December 2025 alongside a base version and a Nemotron 3 white paper. The company also published the full training recipe and the datasets used.

Inside the model

The network has 52 layers in a hybrid arrangement: 23 Mamba-2 layers (a state-space design that tracks a running summary rather than attending over every past token), 23 mixture-of-experts layers, and 6 grouped query attention layers with 2 groups. Each expert layer carries 128 routed experts plus one shared expert, and 6 experts are switched on per token. Training used 25 trillion tokens with a batch size of 3072 and a warmup-stable-decay learning rate schedule. Pre-training drew on crawled and synthetic code, maths, science and general knowledge data, run with Megatron-LM. Later stages added supervised fine-tuning, then multi-environment reinforcement learning with synchronous GRPO, plus RLHF guided by a generative reward model.

What it can do

NVIDIA reports scores across knowledge, maths, coding and agent benchmarks. In its own measurements the model reached 78.3 on MMLU-Pro, 89.1 on AIME25 without tools and 99.2 with tools, 68.3 on LiveCodeBench v6, and 38.8 on SWE-Bench with OpenHands. It scored 50.0 on MiniF2F pass@1 and 79.9 at pass@32. Long-context RULER results were 92.9 at 256k tokens, 91.3 at 512k and 86.3 at 1M. Turning off the reasoning trace lowers accuracy on harder prompts. Multilingual and long-context scores trail the Qwen3 comparison model in several cases, such as MMLU-ProX.

How to use it

The weights are downloadable from Hugging Face under the NVIDIA Nemotron Open Model License, and NVIDIA says the model is ready for commercial use. Nemotron is described as a family of open models with open weights, training data and recipes. A base checkpoint is published separately, and the training recipe, datasets and evaluation configs are released through NVIDIA's NeMo repositories.

Safety and testing

NVIDIA publishes the training, testing and evaluation datasets and provides a reproducibility tutorial with all evaluation configs in the NeMo Evaluator SDK. Evaluations were run with the Nemo Evaluator SDK and Nemo Skills containers. No separate red-teaming or safety incident details are given here.

Reception and debate

The company positions the model as efficient and accurate for specialised AI agents, and sets its benchmark numbers against Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B. It leads those comparisons on several reasoning, coding and instruction-following tests, while trailing on some knowledge, multilingual and creative-writing measures.

Why it matters

As part of the Nemotron line, the model extends NVIDIA's open family aimed at building specialised AI agents, and it is offered with its training recipe, datasets and evaluation tooling so others can reproduce and adapt it. The design combines Mamba-2, attention and experts to cut the active parameter count for a 30B model.

Questions people ask about Nemotron 3 Nano

Who made Nemotron 3 Nano and when did it come out?

NVIDIA Corporation trained it from scratch and released it on Hugging Face on 15 December 2025. Work ran from September to December 2025. It belongs to the Nemotron family of open models, following the Nemotron 2 line.

How big is it and why does only part of it run at once?

It holds 30 billion parameters in total, but a mixture-of-experts design activates just 6 of 128 routed experts per token, so only about 3.5 billion parameters are used for each token. That keeps it cheaper to run while retaining capacity.

What is unusual about its architecture?

It is a hybrid. Of its 52 layers, 23 are Mamba-2 state-space layers, 23 are mixture-of-experts layers, and 6 use grouped query attention. NVIDIA calls the result a Mamba2-Transformer Hybrid MoE, aimed at efficiency in agent and reasoning work.

Can I use it commercially?

Yes. NVIDIA says the model is ready for commercial use, under the NVIDIA Nemotron Open Model License. Weights, training data and recipes are published openly as part of the Nemotron family.

Which languages does it support?

English is the main target, along with coding languages. NVIDIA also lists German, Spanish, French, Italian and Japanese as supported, and reports translation and multilingual benchmark results for them.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Nemotron 3 Nano yet, so the facts on this page come from the sources listed above: huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.