Nvidia Unveils Nemotron 3, Its AI Agent Model Family

Nvidia has unveiled Nemotron 3, a new family of open language models designed to build AI systems in which multiple agents collaborate. The first model, Nemotron 3 Nano, is available to download and run now; the larger Super and Ultra versions are due in the first half of 2026.

The announcement matters because Nvidia is no longer just selling the chips used to train and run models. With Nemotron, the company also wants to provide a software and data foundation for developers building assistants, automations and specialized agents inside businesses.

A 30-billion-parameter model that doesn’t activate everything at once

Nemotron 3 Nano has 30 billion parameters, the internal values a model learns during training. But it activates up to 3 billion of them to process each token, the small unit of text these systems work with.

That difference comes from a mixture-of-experts architecture, or MoE. Rather than mobilizing the entire neural network for every question, the model selects the components best suited to the task. The goal is to preserve reasoning capability without paying the computational cost of a dense 30-billion-parameter model for every response.

Nvidia calls its design a hybrid mixture-of-experts architecture and says Nano delivers up to four times the token throughput of Nemotron 2 Nano. It also says the model reduces the number of tokens generated during reasoning by up to 60%—a significant figure because those internal steps typically increase both latency and the inference bill: the cost of using an already-trained model.

The model includes a 1-million-token context window. In practical terms, it can handle extensive document collections, lengthy histories or detailed instructions without losing the initial information as quickly. That does not, by itself, guarantee correct answers, but it does expand the amount of material the model can consider in a task.

Artificial Analysis, an independent organization that benchmarks models, has ranked Nemotron 3 Nano among the most efficient open models of comparable size, with notable accuracy.

From a chatbot to a team of agents

Nvidia is positioning Nemotron 3 around a trend gaining ground in business: replacing a single assistant with a group of agents. In that setup, one agent might search for information, another write code, a third review the results and another decide which task comes next.

The approach offers advantages for complex processes, but it also introduces problems of its own. Agents must exchange information, maintain context and avoid duplicating work. If every step relies on a large, expensive model, the system can become slow or difficult to sustain economically.

Nemotron 3 Nano is designed for tasks such as content summarization, information retrieval, workflow assistance and software debugging. Nvidia’s strategy is for developers to reserve proprietary frontier models for tasks that genuinely require them and use open, customizable and cheaper models for everything else.

It is a pragmatic approach: the immediate future of enterprise agents does not appear to be an absolute choice between open and closed models, but systems that route each task to the model best suited to it.

Super and Ultra raise the stakes for 2026

The family will have two higher-end tiers. Nemotron 3 Super will have roughly 100 billion parameters and activate up to 10 billion per token. Nemotron 3 Ultra will reach approximately 500 billion parameters, with up to 50 billion active per token, and will target more demanding reasoning, research and planning tasks.

Both models will use Nvidia’s 4-bit NVFP4 numerical format on the company’s Blackwell architecture. Working at lower numerical precision reduces the memory required to train and run models, although it calls for careful techniques to prevent the loss of precision from degrading their results.

For now, Nano is the actual product. Super and Ultra define the roadmap and will need to be evaluated when they become available—not just by their size, but by their performance on real-world tasks, hardware requirements and how easily they can be tuned to proprietary data.

Models, data and tools in one package

Nvidia is supporting the launch with three trillion tokens of data for pretraining, post-training and reinforcement learning. The latter method trains a model through rewards for completing a task successfully, and is especially useful for improving its ability to follow multistep processes.

The company has also released the NeMo Gym and NeMo RL libraries for creating training environments, as well as NeMo Evaluator for checking performance and safety. The resources are available on GitHub and Hugging Face, along with a dataset focused on evaluating the safety of agentic systems.

Nano is compatible with popular tools for running models locally or on servers, including LM Studio, llama.cpp, SGLang and vLLM. It will also be available through inference providers and as an Nvidia NIM microservice for deployments on the company’s accelerated infrastructure.

The move reinforces a clear ambition: that choosing Nvidia hardware should also make it easier to choose the company’s models, training tools and production infrastructure. For developers and businesses, the value of the ecosystem will come down to whether it offers an open alternative with enough performance, control and predictable costs to compete with the closed models that currently dominate many agent applications.