Nemotron 3 is NVIDIA’s new open model family for agentic AI systems where multiple cooperating agents plan, retrieve, verify, and use tools over long contexts and long-running workflows.
-
Built for agentic workflows: Designed for retrievers, planners, tool executors, and verifiers operating together over large contexts and long horizons.
-
Hybrid Mamba–Transformer MoE: Combines Mamba layers, transformer layers, and Mixture-of-Experts for efficient, long-range reasoning and high throughput.
-
1M-token context window: Enables sustained reasoning across large codebases, document collections, extended conversations, and multi-step plans.
-
Reinforcement learning in NeMo Gym: Trained with multi-environment RL for reliable multi-step behavior, tool use, and trajectory-style planning.
-
Open and transparent: Model weights, training data sources, and recipes are openly released so developers can inspect, customize, and extend Nemotron 3.
Nemotron 3 launches with three variants—Nano, Super, and Ultra—with Nemotron 3 Nano available today, plus ready-to-use cookbooks for major inference engines to help you get up and running quickly.
📚 Check out these resources:
- NVIDIA Technical Report: https://nvda.ws/44roJV6
- NVIDIA Whitepaper: https://nvda.ws/4iTeLSd
- Hugging Face Model Page: https://nvda.ws/4pvrEEA
👨🍳 Deployment Cookbooks:
vLLM Cookbook: https://nvda.ws/48WNmu9
SGLang Cookbook: https://nvda.ws/4oONkKD
TRT-LLM Cookbook: https://nvda.ws/44t8zum
Tech Blog: https://nvda.ws/4rWXJ9Q
Stay up to date on NVIDIA Nemotron by subscribing to NVIDIA news and following NVIDIA AI on LinkedIn, X, YouTube, and the Nemotron channel on Discord.
*Access open Nemotron Models on Hugging Face and a collection of NIM microservices and Developer Examples on build.nvidia.com. Share your ideas and vote on what matters to help shape the future of Nemotron. *