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 Nano 2

Nemotron Nano 2

A 9B hybrid Mamba-Transformer model for reasoning and chat, from NVIDIA.

Nemotron Nano 2, released as NVIDIA-Nemotron-Nano-9B-v2, is a hybrid Mamba-Transformer language model trained from scratch by NVIDIA for both reasoning and non-reasoning work. It came out on 18 August 2025 and is offered for commercial use under the NVIDIA Open Model License.

Share on WhatsApp

  • Hybrid design: mostly Mamba-2 and MLP layers, just four attention layers
  • Pretrained as a 12B model on 20 trillion tokens, then compressed to 9B
  • Reasoning can be toggled on or off via system prompt, with a thinking budget
  • Context window up to 128K tokens
  • Released 18 August 2025 under the NVIDIA Open Model License

What it is

NVIDIA-Nemotron-Nano-9B-v2 is a large language model built from scratch by NVIDIA. It handles both reasoning tasks and ordinary ones in a single model: it writes out a reasoning trace first, then gives a final answer. A system prompt controls whether that trace appears at all. Turning it off is possible, though NVIDIA notes a small accuracy drop on harder prompts. The model has 9 billion parameters and a context window of up to 128K tokens. NVIDIA positions it for chatbots, agents, retrieval-augmented generation and instruction following, mainly in English and coding languages. It was released on 18 August 2025 and is cleared for commercial use.

How it came about

The model was trained from scratch at NVIDIA and published on 18 August 2025, both on Hugging Face and in the company's API catalogue. It builds on the Nemotron-H architecture described in an earlier technical report. A companion paper, listing more than 200 NVIDIA authors, was posted to arXiv on 20 August 2025 and revised on 2 September 2025. That paper describes how a 12-billion-parameter base model was trained first, then compressed to the released 9B version.

Inside the model

The network is a Nemotron-Hybrid design: mostly Mamba-2 and MLP layers, with only four attention layers. Mamba-2 is a sequence model that keeps a running state instead of comparing every token with every other, which speeds up generation of long thinking traces. Training started from Nemotron-Nano-12B-v2-Base, pretrained on 20 trillion tokens with an FP8 recipe, then aligned and compressed using the Minitron strategy. NVIDIA says the model was improved using Qwen. Pretraining data has a September 2024 cutoff. Training used Megatron-LM and NeMo-RL.

What it can do

NVIDIA reports benchmark results with reasoning switched on, except RULER which was run with reasoning off. On IFEval (Instruction Strict) the model scored 90.3% against 89.4% for Qwen3-8B. On RULER at 128K it scored 78.9% against 74.1% for the same model. Evaluations were run with NeMo-Skills and NVIDIA published a tutorial for reproducing them. Supported languages are listed inconsistently: one section names English, German, Spanish, French, Italian and Japanese, while the input specification also lists Korean, Portuguese, Russian and Chinese.

How to use it

Weights are downloadable from Hugging Face and served through NVIDIA's API catalogue. Use is governed by the NVIDIA Open Model License Agreement, and NVIDIA states the model is ready for commercial use. It runs in Hugging Face Transformers, TensorRT-LLM and vLLM, and is supported on NVIDIA A10G, A100, H100-80GB and Jetson AGX Thor hardware.

Safety and testing

NVIDIA gives no safety evaluations, red-teaming results or incident reports for this model in the available material. The company does note that its models are optimised for NVIDIA GPU-accelerated systems, and warns that vLLM deployments need a specific cache setting to avoid degraded accuracy.

Reception and debate

No independent reviews, criticism, lawsuits or bans are described in the available material. NVIDIA presents the model as accurate and efficient for reasoning workloads relative to similarly sized models, and runs a public feature-voting board for user requests about the Nemotron line.

Why it matters

The model is presented as an efficient alternative for reasoning workloads, using Mamba-2 layers to cut the cost of long thinking traces. Its design follows the Nemotron-H architecture, and the paper describes a compression path from a 12B base model to the released 9B version using the Minitron strategy.

Questions people ask about Nemotron Nano 2

What is Nemotron Nano 2?

It is a 9-billion-parameter language model from NVIDIA, released in August 2025. It combines a Mamba-2 sequence model with a small number of attention layers and is built for both reasoning tasks and ordinary chat, with a context window of up to 128K tokens.

How does the reasoning control work?

A system prompt decides whether the model shows its reasoning. Sending /think, or nothing, turns reasoning on; /no_think turns it off. Developers can also cap how many tokens the model spends thinking, which NVIDIA says helps meet response-time targets on agents and edge devices.

How can I run it?

The weights are on Hugging Face and the model is also served through NVIDIA's API catalogue. It works with Hugging Face Transformers, TensorRT-LLM and vLLM, and NVIDIA lists support for its A10G, A100, H100-80GB and Jetson AGX Thor hardware.

What licence does it use?

Use is governed by the NVIDIA Open Model License Agreement. NVIDIA states the model is ready for commercial use.

Which languages does it support?

NVIDIA lists English, German, Spanish, French, Italian and Japanese in one place, and adds Korean, Portuguese, Russian and Chinese in the input specification. English and coding languages are the primary focus.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Nemotron Nano 2 yet, so the facts on this page come from the sources listed above: huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2, arxiv.org/abs/2508.14444. 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.