NVIDIA unveils Nemotron 3 to supercharge AI agents

Nemotron 3 is NVIDIA's new family of open models tailored specifically for multi-agent AI systems, designed to make complex agentic workflows both accurate and cost-effective.

Running fleets of interacting AI agents usually triggers huge compute bills, communication overhead, and context drift. Nemotron 3 tackles this by adopting a hybrid latent mixture-of-experts (MoE) architecture that keeps active parameter counts exceptionally low during inference without sacrificing reasoning accuracy.

Three sizes built on sparse MoE

The lineup spans three distinct tiers designed for different workloads.

Leading the release is Nemotron 3 Nano, which is available today. It packs 30 billion total parameters while activating only up to 3 billion per token. NVIDIA reports that Nano delivers 4x higher token throughput than Nemotron 2 Nano and slashes reasoning token generation by up to 60%. It also comes with a massive 1-million-token context window. Independent benchmarking firm Artificial Analysis ranked it as the most open and efficient model in its class.

Two larger variants are scheduled for the first half of 2026. Nemotron 3 Super features 100 billion parameters with 10 billion active, targeting latency-sensitive multi-agent collaboration. The flagship Nemotron 3 Ultra scales to 500 billion parameters with 50 billion active for heavy research and strategic planning. Both larger models utilize NVIDIA's 4-bit NVFP4 training format on Blackwell infrastructure, keeping memory footprints low during distributed training.

Complete training stacks and open recipes

NVIDIA is releasing far more than raw model weights this time around.

The launch includes 3 trillion tokens of pre-training, post-training, and reinforcement learning datasets, alongside an agentic safety dataset. Developers also get access to NeMo Gym and NeMo RL, two open libraries that provide ready-to-run reinforcement learning environments, which have already been integrated by tooling providers like Unsloth and Prime Intellect.

Crucially, NVIDIA also published NeMo Evaluator alongside full evaluation recipes. Instead of posting vanity benchmark numbers, the company released the exact prompts, harness configurations, and runtime settings so engineers can reproduce the reported benchmarks locally.

For local setups, Nemotron 3 works out of the box with llama.cpp, vLLM, LM Studio, and SGLang. You can grab the Nano weights directly from Hugging Face right now, spin it up via NVIDIA NIM microservices, or query hosted endpoints on platforms like OpenRouter, Fireworks, DeepInfra, and Baseten.

Why it matters

If you build agentic pipelines, tool-calling automations, or local coding assistants, Nemotron 3 Nano is an immediate download candidate. Activating just 3 billion parameters combined with a million-token context window gives you an ultra-cheap, highly responsive workhorse for chaining multi-agent loops without melting your inference budget.

We think NVIDIA is playing a smart long game here. By open-sourcing 3 trillion tokens of training data, RL environments, and reproducible evaluation harnesses, it sets a much higher transparency bar for the open-weight ecosystem while ensuring developers build their multi-agent stacks directly on top of NVIDIA's software fabric.