NVIDIA launches Nemotron 3 Nano, an open AI model

NVIDIA has introduced Nemotron 3 Nano, an open artificial intelligence model designed to run agents and multi-step workflows without driving up costs or response times.

The model combines 31.6 billion parameters, although it activates only around 3.6 billion per token. A parameter is an internal unit the model adjusts during training, while a token is a fragment of text it processes. In practice, this gives it the capabilities of a large model with usage closer to that of a small one.

A model built to work as an agent

Nemotron 3 Nano is designed for tasks where AI does more than respond. It also plans, uses tools and completes multi-step processes. For example, it can search for information, analyze it, write code and use an external application within the same workflow.

It uses a hybrid architecture that combines three approaches:

• Mamba-2, designed to process long conversations and documents with low latency.
• A Transformer architecture, which is more precise when focusing on specific details and solving complex problems.
• A mixture-of-experts system, or MoE, that activates only part of the model for each operation.

The system has 128 internal experts and activates six for each step. They do not all work at the same time. This selection reduces the work required without removing the total capacity available.

NVIDIA says the model can reach up to four times the speed of Nemotron Nano 2 and up to 3.3 times the speed of other models in its category, although those figures depend on the hardware and configuration used. In a test with an H200 GPU, 8,000 input tokens and 16,000 output tokens, it achieved 3.3 times the performance of Qwen3-30B and 2.2 times that of GPT-OSS-20B.

One million tokens of context

The context window extends to 1 million tokens. This is the amount of information the model can take into account during an interaction, although it does not correspond exactly to one million words.

That limit makes it possible to work with very long documents, complete project histories or large collections of information. It can also provide a broader memory for agents operating over long periods, such as when reviewing hundreds of files or coordinating several related tasks.

The model also offers controls for deciding how much reasoning it performs before responding:

• Reasoning ON, for tasks that need step-by-step analysis.
• Reasoning OFF, for more direct and faster responses.
• A configurable reasoning-token budget to limit the time and cost of each query.

This avoids using the same level of processing for everything. A simple question can receive an immediate answer, while a programming problem can be given more time for analysis.

The open release includes data and tools

NVIDIA is not only publishing the model weights, which are the files needed to run it. It is also releasing much of the data, code and training recipe under the nvidia-open-model-license.

The process included a corpus of 25 trillion tokens, 13 million post-training samples and more than ten reinforcement-learning environments with over 900,000 mathematics, programming, reasoning and tool-use tasks. It also includes nearly 11,000 labeled traces for analyzing security risks in agents.

Alongside the model, the company is introducing NeMo Gym, an open library for creating environments where other models can practice and be evaluated. These environments can simulate everything from mathematics and programming exercises to calendars, multi-turn conversations and external tool use.

The idea is to let developers train their own models with verifiable tasks, collect errors and test changes without having to build the entire infrastructure from scratch.

What this changes for you

If you build AI applications, Nemotron 3 Nano offers an open alternative for running many queries in parallel, especially on GPU servers. Its one-million-token context can reduce the need to split large documents or constantly summarize the history.

It is also available to try through Hugging Face, OpenRouter and NVIDIA’s services. It can be deployed with tools such as vLLM and SGLang, and run on systems with NVIDIA hardware through solutions such as Llama.cpp, LM Studio and Unsloth.

The main question is not only its performance figures, but how it will behave outside NVIDIA’s tests: on real tasks, with specific costs, different GPUs and proprietary data. Even so, the launch points to a clear trend: multi-agent systems need models that are capable enough, but also cheap and fast enough to run through many stages in succession.