Post
335
Nemotron 3 Nano (30B A3B): The Efficient Open Agent Model 🔥
NVIDIA Nemotron 3 Nano is a new standard for high-throughput, accurate AI agents, built on a groundbreaking Hybrid Mamba-Transformer Sparse Mixture-of-Experts (MoE) architecture.
Speed: 4x faster than Nemotron 2 Nano
Scale: 1 Million-token context window.
Intelligence: Best-in-class reasoning accuracy with ~3.6B active parameters per token.
Control: Features Reasoning ON/OFF modes and a customizable Thinking Budget.
Open: Full open weights, training recipes, and all major datasets (including 13M post-training samples and NeMo Gym RL environments).
Get Started: Download the model and datasets on Hugging Face: https://huggingface.co/blog/nvidia/nemotron-3-nano-efficient-open-intelligent-models
NVIDIA Nemotron 3 Nano is a new standard for high-throughput, accurate AI agents, built on a groundbreaking Hybrid Mamba-Transformer Sparse Mixture-of-Experts (MoE) architecture.
Speed: 4x faster than Nemotron 2 Nano
Scale: 1 Million-token context window.
Intelligence: Best-in-class reasoning accuracy with ~3.6B active parameters per token.
Control: Features Reasoning ON/OFF modes and a customizable Thinking Budget.
Open: Full open weights, training recipes, and all major datasets (including 13M post-training samples and NeMo Gym RL environments).
Get Started: Download the model and datasets on Hugging Face: https://huggingface.co/blog/nvidia/nemotron-3-nano-efficient-open-intelligent-models