NVIDIA

Nemotron 3 Nano Omni

One model to see, hear, and reason.

30B-parameter MoE (3B active). Text, images, audio, and video in a single forward pass — plus a meeting intelligence agent and Gradio UI.
NVIDIA Nemotron 3 Nano Omni
Text

Reasoning + tool calling

Images

OCR, charts, UI screens

Audio

ASR, meetings, calls

Video

Converged A/V context

EFFICIENCY
30B capacity · 3B cost

Hybrid mixture-of-experts: deep expertise when you need it, inference priced like a 3B model. Up to 9× more concurrent users on the same GPU.

ONE PASS
Unified perception

No ASR → vision → fusion pipeline. Cross-modal reasoning happens in one representation space — audio can inform what the model sees.

DEVELOPER UX
OpenAI-compatible API

enable_thinking, streaming, tool calls across modalities, and per-request reasoning depth.

Architecture

From fragmented stacks to one model

Fragmented perception stack

Before: four models, context loss at every handoff, latency stacks.

Unified perception stack

Now: one deployment, one latency budget, shared cross-modal context.

Benchmarks
Meeting intelligence

Process meetings in one inference call

The meeting-agent ingests audio, video, images, and text — then extracts transcripts, action items, decisions, and open questions. Includes a NVIDIA-green Gradio UI with every model setting exposed.

Gradio UI
Reasoning toggle Modality cheat sheet
Nemotron 3 model family
Repository

What's inside

blog.md

Deep-dive with architecture analysis and code examples
meeting-agent/

Agent, modalities, demo, and Gradio UI
Colab notebook

Hands-on omni-modal experiments
Blog PDF

Print-ready version of the write-up
Briefing slides

Executive summary deck
images/

Diagrams, benchmarks, and UI screenshots
Quick start

Run the meeting agent locally

git clone https://github.com/cobusgreyling/NVIDIA-Nemotron-3-Nano-Omni.git
cd NVIDIA-Nemotron-3-Nano-Omni/meeting-agent
pip install openai requests Pillow gradio python-dotenv
export NVIDIA_API_KEY="nvapi-..."

# Self-contained demo
python examples/demo.py

# Full Gradio UI
python ui.py
# open http://localhost:7860

Get an API key from build.nvidia.com. Full setup and modality examples are in blog.md.