< 1B – 30B MoE

NVIDIA

The Nemotron 3 family — efficient open models for reasoning, multimodal, speech and document AI.

Website Hugging Face X GitHub
01What they offer

NVIDIA’s Nemotron 3 family spans tiny edge models to MoE reasoning models — all efficient, open, and built to run on hardware you can actually get. Mix and match across reasoning, multimodal, speech and document extraction.

02Models & products

The kit, sized small

Nemotron 3 Nano

30B total · 3B active MoE — efficient reasoning for long-running agents.

  • MoE
  • reasoning
30B parameters32B cap
Nemotron 3 Nano 4B

Edge-optimised 4B model for constrained hardware.

  • edge
  • text
4B parameters32B cap
Nemotron 3 Nano Omni

Multimodal nano model across modalities.

  • omni
  • multimodal
Nemotron 3 ASR

Speech recognition built for real-time use.

  • speech
  • ASR
Nemotron Parse

Sub-1B parameter document extraction.

  • documents
  • extraction
1B parameters32B cap
Nemotron Embed VL

Vision-language embeddings for retrieval & search.

  • embeddings
  • vision
03Build this with us

If you want to build…

A long-running agent / reasoning app
Nemotron 3 Nano3B active MoE
To run on the edge
Nemotron 3 Nano 4BEdge-ready
A multimodal app
Nemotron 3 Nano OmniOmni-modal
Speech recognition
Nemotron 3 ASRReal-time ASR
Document extraction
Nemotron Parse< 1B params
04Getting started
Nemotron 3 Nano usage guide Nemotron 3 Ultra blog Introducing NVIDIA Nemotron 3 Nano Omni NVIDIA Nemotron collections
STARTER SPACE
Nemotron-3 Nano Omni demo

Fork this Gradio Server Space to start from a working Nemotron-3 Nano Omni app.

Fork it
05Eligible prize
NVIDIA · SPONSOR PRIZE

Nemotron Hardware Prize

To qualify ·Build with Nemotron models. Awarded for best Space (judged) and community engagement.

2× RTX 5080
All prizes