🟩 Nemotron 3 Nano Launch, what you need to know

Nemotron 3 Nano is an open 30B model built for agent workflows. It mixes Mamba‑2 and Transformer layers with a Mixture‑of‑Experts router that activates a few experts per token (6 of 128), so you get stronger reasoning without running all 31.6B params every step (~3.6B active). (aka longer context while staying cheap).

Why it matters

• - It’s fast: up to 4x higher token throughput than Nemotron 2 Nano, and strong throughput versus similar open models in like‑for‑like tests.

• - It handles long input: up to 1M tokens. Most people will use smaller windows due to VRAM, but training included a 512k stage to keep short‑task accuracy steady.

• - It’s controllable: toggle reasoning on/off and set a thinking budget to cap “reasoning” tokens. Helps keep multi‑agent costs predictable.

• 
  

Quality and training

• - Good results in math and coding benchmarks (GSM8K, MATH, HumanEval, MBPP), plus solid long‑context scores as seen in the scores for the RULER benchmark below, with scores from Nemotron 3 Nano highlighted on the right.

• - Post‑training uses supervised fine‑tuning, reinforcement learning across multiple environments, and RLHF. NeMo Gym and - NeMo RL are open so you can train and evaluate in similar setups.

• 
  

  

• 
  

• 
  

Open and deployable

• - Open weights, recipes, and large slices of datasets are released. Licensed for commercial use under NVIDIA’s Open Model License, with derivatives allowed and no claim on outputs. (HuggingFace)

• - Run it on H100/A100 with vLLM or SGLang, or use llama.cpp/LM Studio for local tests. It’s available via Hugging Face and major inference providers.

• 
  

TL;DR

• If you’re building agents that need speed, long context, and cost controls, Nemotron 3 Nano is a good starter. Route harder tasks to a frontier model when needed, and keep routine work on Nano.

• Nemotron 3 Super (~100B) and Ultra (~500B) are planned for early 2026, aimed at higher‑end reasoning while keeping efficiency in mind, as per their news release.
  
  Read more on our in-depth analysis: https://llm-stats.com/blog/research/nemotron-3-nano-launch

• Use today through our playground: https://llm-stats.com/playground