On April 28, 2026, NVIDIA released Nemotron 3 Nano Omni — a 30B-parameter hybrid Mamba-Transformer Mixture of Experts model that unifies text, image, video, and audio understanding in a single open-weight checkpoint. Details from the NVIDIA Blog announcement and the NVIDIA Technical Blog deep dive. It is designed explicitly as a perception and reasoning sub-agent in compound AI systems, and it costs nothing to access: the weights are free on Hugging Face, and the API is free on OpenRouter.
The technical headline is unusual: ~31 billion total parameters, but only ~3 billion active per forward pass, per the Hugging Face model card. That active-parameter count is what determines inference cost.
This guide covers the architecture, benchmarks, deployment options, and agentic use cases for builders evaluating Nano Omni as a component in multi-agent systems.
The One-Sentence Summary
Nemotron 3 Nano Omni is a ~30B-total/~3B-active Mamba-Transformer MoE model that accepts text, images, video, and audio as input, scores 82.1% on AIME 2025 (no tools) and 63.2% on LiveCodeBench in its own technical report — a modest step down from its text-only sibling Nemotron 3 Nano (89.1% / 68.3% on the same two benchmarks), the expected tradeoff of unifying reasoning across four modalities — and is available as open weights and free API, built specifically to serve as the multimodal perception layer in compound agent architectures.
Why This Matters for Builders
Three constraints have historically made omnimodal agents expensive or impractical:
| Problem | Previous workaround | Nano Omni approach |
|---|---|---|
| Separate model stacks per modality | Run vision model + audio model + LLM in sequence | Single model handles all modalities in one reasoning pass |
| Omnimodal = slow | Accept high latency or sacrifice quality | Mamba layers eliminate the quadratic attention bottleneck |
| Frontier capability = large active footprint | Use a 70B+ dense model or pay closed-model API rates | MoE routing keeps active parameters at ~3B |
If you are building a system that needs to reason across a video transcript, an uploaded image, a voice memo, and a text document in the same context window — and you want to do this affordably, with open weights, on-premise — Nano Omni is one of the few models that addresses all four constraints at once (we have not surveyed every open omnimodal release to confirm it is literally the first).
Architecture: Mamba-Transformer Hybrid MoE
Why Mamba?
Standard transformer attention scales quadratically with sequence length. For an omnimodal model processing a video (many frames) alongside a long document, this quadratic cost becomes prohibitive. Mamba-2 is a state-space model that replaces the attention mechanism with a recurrent formulation, achieving linear sequence complexity — the cost of processing frame 1000 of a video is the same as processing frame 1.
NVIDIA’s design combines Mamba’s sequence efficiency with transformer attention’s ability to perform precise local reasoning across a context window. The result is a hybrid:
| Layer type | Count | Role |
|---|---|---|
| Mamba-2 layers | 23 | Sequence compression, memory efficiency, long-range state |
| MoE Transformer layers | 23 | Sparse expert routing for domain specialization |
| Attention layers (GQA) | 6 | Precise local context reasoning |
Total: 52 layers. Model dimension: 2688. Query heads: 32. KV heads: 2. Head dimension: 128. (Per the model’s published config.json.)
Why MoE?
The ~31B total parameters are divided among 128 routed experts per MoE layer plus 1 shared expert. Each forward pass activates 6 of the 128 routed experts per token (the shared expert activates on every token) — confirmed in the model config. The result: ~3 billion parameters active per inference call — a compute cost comparable to a small dense model, with capacity comparable to a much larger one.
The practical implication: Nano Omni can run inference on a single GPU rather than requiring multi-GPU tensor parallelism — a single H100 80GB at full BF16 precision, or as small as a single RTX 5090 32GB using the NVFP4-quantized release, per the Hugging Face model card. This matters for edge agents and on-premise deployments where multi-node infrastructure is not available.
Omnimodal: Native, Not Adapted
Most multimodal models attach vision or audio via an adapter trained on a frozen language backbone. Nano Omni integrates all modalities from training step zero — there is no seam between the vision encoder, audio processor, and language reasoning. This matters for tasks that require cross-modal reasoning: interpreting a chart from a screenshot while writing code to reproduce it, or summarizing a video while citing timestamps in a document.
Accepted inputs (per the Hugging Face model card):
- Text (up to a 256K-token context window)
- Images (static)
- Video (multi-frame)
- Audio
Output: Text only.
Benchmarks in Context
Math and Reasoning
NVIDIA’s own Nano Omni technical report (Table 10) benchmarks the multimodal model’s text reasoning directly against its text-only sibling, Nemotron 3 Nano:
| Benchmark | Nano Omni | Nemotron 3 Nano (text-only sibling) |
|---|---|---|
| AIME 2025 (no tools) | 82.1% | 89.1% |
| GPQA (no tools) | 72.2% | — |
| MMLU-Pro | 77.3% | — |
Unifying text, image, video, and audio reasoning into a single model costs Nano Omni roughly 7 points of AIME accuracy versus the text-only sibling that shares its backbone — a real tradeoff, not a wash. NVIDIA’s technical report does not publish a Nano-Omni-specific comparison against Qwen3-30B-A3B or GPT-OSS-20B on these text benchmarks, so this guide does not claim one.
Code
| Benchmark | Nano Omni | Nemotron 3 Nano (text-only sibling) |
|---|---|---|
| LiveCodeBench | 63.2% | 68.3% |
Same pattern as AIME: the omnimodal model trails its text-only sibling on LiveCodeBench, per the same technical report table. No head-to-head Nano-Omni-vs-Qwen3-30B-A3B or Nano-Omni-vs-GPT-OSS-20B code benchmark has been published by NVIDIA as of this writing.
Throughput and Multimodal
MediaPerf — an open benchmark from Coactive that scores multimodal models on real media-industry video tasks, reporting quality, cost, and latency together — found Nano Omni delivering the highest throughput and lowest inference cost of any benchmarked model, open or closed, on the video-tagging task (9.91 hours of video processed per hour, at $14.27 per unit of work), beating GPT-5.1, Gemini 3.0 Pro, and Qwen3-VL as well as other open omnimodal models. NVIDIA’s own technical report cites, specifically versus Qwen3-Omni on a B200 GPU: up to 3x higher single-stream output token throughput, and up to 9x higher output token throughput per GPU at a fixed interactivity target.
These numbers reflect the architectural benefit of Mamba-2 layers: processing long video frame sequences does not incur the quadratic attention penalty that hits transformer-only competitors. (Note: a commonly repeated “3.3x faster than GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507” throughput figure comes from NVIDIA’s report on the text-only Nemotron 3 Nano sibling, not from Nano Omni — it does not apply to the multimodal model covered here.)
Where Sub-Agent Design Fits
NVIDIA explicitly positions Nano Omni as a perception and context sub-agent, not an orchestrator. In compound agent architectures, the model makes sense as the layer that:
- Ingests raw multimodal input (a user-uploaded PDF, a screen recording, a voice note) and converts it to structured text context
- Performs initial analysis (summarize, extract entities, identify anomalies)
- Passes context to an orchestrating model (Claude, GPT-5.x, or similar) for decision-making and action planning
This fits a common pattern in enterprise AI systems where a specialized, efficient perception model runs locally or at low cost, and the more expensive orchestrating model only receives processed context — not raw frames or audio.
Example architecture:
[Raw input: video + document + voice memo]
↓
[Nano Omni — multimodal perception sub-agent]
↓ structured summary, extracted facts
[Orchestrator — Claude / GPT-5.x / internal LLM]
↓ decisions + tool calls
[Action layer — APIs, databases, downstream tools]
The benefit: Nano Omni’s per-token cost is near zero (free API tiers exist), so you can process large volumes of raw multimodal input without expensive orchestrator tokens.
Deployment Options
Free API Access
| Platform | Access | Notes |
|---|---|---|
| OpenRouter | Free tier | nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free |
| build.nvidia.com | NVIDIA NIM | Hosted API endpoints are free for prototyping; production deployment requires an NVIDIA AI Enterprise license (a free 90-day evaluation license is available) |
| fal.ai | Available | Serverless inference, per the NVIDIA technical blog |
Open Weights (Self-Hosted)
Full weights are available on Hugging Face under NVIDIA’s open model license. Review the model card for commercial use terms before deploying in a production product.
Supported inference frameworks, per the Hugging Face model card:
- vLLM — recommended for high-throughput server deployments
- SGLang — structured generation, tool-calling pipelines
- Ollama — local desktop or developer-workstation deployment
- llama.cpp — CPU-capable; useful for true edge devices
NVIDIA NIM (Enterprise)
For production enterprise deployments, NVIDIA offers Nano Omni as a NIM microservice. Production use requires an NVIDIA AI Enterprise license. Per the NVIDIA launch blog, early adopters include Palantir and Foxconn (among others); Dell Technologies is listed as evaluating the model, not yet a confirmed production adopter.
When to Use Nano Omni vs. Alternatives
| Scenario | Recommendation |
|---|---|
| Need multimodal perception at zero API cost | Nano Omni — free on OpenRouter and Hugging Face |
| Need to process video + audio + images in one model | Nano Omni — unifies all four modalities in one open-weight checkpoint at ~3B active parameters |
| Running a perception sub-agent at edge (single GPU) | Nano Omni — 3B active parameters, supports Ollama/llama.cpp |
| Need proprietary fine-tuning on domain video/audio | Nano Omni — open weights allow this |
| Need maximum raw coding or reasoning performance | Evaluate closed models (Claude, GPT-5.x); Nano Omni is strong but not at closed-frontier ceiling |
| Text-only agent with no multimodal input | Nano Omni works but Qwen3-30B-A3B or similar may offer better text-only throughput/cost tradeoffs |
| Long-context document processing (256K+) | Nano Omni supports 256K; for beyond that, check provider-specific limits |
What to Watch
- Community fine-tunes: The open weights will attract domain-specific variants — medical imaging sub-agents, industrial video analysis models, audio transcription specialists built on Nano Omni’s base.
- Self-hosting throughput data: Real-world vLLM benchmarks on A100/H100 will determine whether 9x throughput advantage holds on user hardware vs. NVIDIA’s reference clusters.
- License clarification: NVIDIA’s open model licenses have historically included restrictions (commercial use, redistribution). Verify the current Hugging Face model card before committing to a production build.
- Competitor response: At the 30B-A3B efficiency class, Qwen3-30B-A3B-Thinking and GPT-OSS-20B are the nearest text-only competitors. A true omnimodal equivalent from either family would change the landscape.
This analysis is based on NVIDIA’s published technical blog, the Nemotron 3 Nano technical report (arXiv), OpenRouter and DeepInfra benchmark data, and the Hugging Face model card as of April–June 2026. ChatForest researches and analyzes public sources — we do not run our own model evaluations or production deployments.