# Nemotron 3 Nano Omni 30B A3B Reasoning (slim)

> Nemotron 3 Nano Omni is NVIDIA's open-weight multimodal model that takes video, audio, images and text and returns text, a hybrid Mamba-Transformer mixture of experts with about 31B parameters and 3B active, published in BF16, FP8 and NVFP4.

- Full: https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning.md (~13,300 tokens) · this version ~2,480 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning.json · canonical https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 60.7/100 · rank #52 of 100 open-weight models · #45 in Open-weight models · not agent-ready · confidence high**


Assessment: A local multimodal cousin for video with speech, long documents, charts, screens and English transcription, with tool calling, a reasoning switch and a 262,144-token context. It needs about 27 GB for NVIDIA's NVFP4 build on Blackwell, or 42 GB for FP8. The main risks are the remote code it loads and a repository where nobody answers questions.

## Facts

- Kind: Open-weight model · vendor: NVIDIA · category: Open-weight models · legal entity: NVIDIA Corporation · provenance 73/100
- Packages: huggingface `nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16`, huggingface `nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8`, huggingface `nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4`, ollama `nemotron3`
- Auth: None · pricing: Free · x402: no · licence: NVIDIA Open Model Agreement (custom, commercial use allowed, not an OSI licence)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 and its FP8 and NVFP4 repositories, read from the model card and its four subcards, the files and their metadata, the chat template, the licence, the technical report, NVIDIA's vLLM and SGLang cookbooks and the runtimes' own code and libraries on 21 October 2026. Nothing was downloaded or run. NVIDIA's hosted API serves the same model and isn't listed here
- Model: The Nemotron 3 Nano 30B A3B hybrid Mamba2-Transformer mixture of experts as the language backbone, with a C-RADIOv4-H vision encoder and a Parakeet speech encoder. Video, audio, images and text in, text out. The card gives 31B parameters with about 3B active; the Hub's safetensors count is 33,015,632,214
- Inputs: Video as mp4 up to 2 minutes (1 fps, 128 frames at 1080p), audio as wav or mp3 up to 1 hour at 8 kHz or more, RGB images and text. English only. Speaker diarisation, translation, multilingual speech and streaming audio or video aren't supported, per the explainability subcard
- Files: Safetensors shards in all three repositories, 66.1 GB (BF16), 35.2 GB (FP8) and 22.4 GB (NVFP4) by the Hub's file lists, with model, processor and configuration code that loads with trust_remote_code. No pickle files. Hugging Face lists a SHA-256 for every shard
- Licence: NVIDIA Open Model Agreement (release date 2 April 2026, footer v. 9 March 2026). Perpetual, royalty-free and irrevocable, commercial use and derivatives allowed, with a patent and copyright litigation termination clause, attribution in a NOTICE file and export compliance. No use policy and no threshold. The card links it from its metadata; the repository has no LICENSE file
- Runtimes: vLLM 0.20.0 or later (NemotronH_Nano_Omni_Reasoning_V3 in its registry), SGLang 0.5.22 (released model code; the card points to dev images), llama.cpp (the nemotron_omni projector, in release b11772), Ollama's library (nemotron3) and LM Studio's catalogue (nvidia/nemotron-3-nano-omni). Transformers loads it through the repository's remote code. NVIDIA also names TensorRT-LLM and TensorRT Edge-LLM
- Context: 262,144 tokens, trained in the sixth SFT stage per the technical report and set as max_position_embeddings in the language config. The card's examples serve at 131,072
- Tools and reasoning: The chat template declares tools inside tools tags and expects calls as function and parameter blocks inside tool_call, which vLLM's and SGLang's qwen3_coder parser reads, and it renders several calls in one turn. NVIDIA's vLLM and SGLang cookbooks declare tools with the OpenAI schema and read tool_calls. Thinking is on by default, enable_thinking false turns it off, and reasoning_budget with vLLM's thinking_token_budget caps it; the nemotron_v3 parser returns the reasoning apart from the answer
- Hardware: The card's minimums are one H100 80 GB for BF16, one L40S 48 GB for FP8 and one RTX 5090 32 GB for NVFP4, with DGX Spark and Jetson Thor also supported. NVFP4 needs NVIDIA Blackwell
- Training data: Listed on the card, 1,395 dataset entries and about 717B tokens by modality, with public, private and NVIDIA-made synthetic sets named with sample and token counts, a 2019 to 2025 collection period, CSAM scanning and two safety judge models for filtering. One training set, Nemotron-Image-Training-v3, is published on Hugging Face

- Scores: Capability 34, Reach 85, Risk 60, Agent interface 90, Documentation and upkeep 65 · total over the 7 assessed categories
- Why: Capability, From 8 independent benchmark results in 4 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns 20 because the smallest 4-bit builds, Ollama's NVFP4 at 21 GB and NVIDIA's NVFP4 at 22.4 GB, need about 25 to 27 GB with the fi… · Risk, File safety gets 15, safetensors only but loaded with trust_remote_code. · Agent interface, A Jinja template ships in the repository (20). · Documentation and upkeep, The card covers all seven fields, reading the 2019 to 2025 data collection period as the cut-off (20).
- Sources: 24, open questions: 5, both in the full twin
- Capabilities: speech.stt, speech.batch, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-nano-omni-30b-a3b-reasoning.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-nano-omni-30b-a3b-reasoning.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-nano-omni-30b-a3b-reasoning from a page on nvidia.com or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Start it with `vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 --max-model-len 131072 --trust-remote-code --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder --kv-cache-dtype fp8` on vLLM 0.20.0 or later, after `pip install "vllm[audio]"` if any audio will be sent
2. Or run `ollama run nemotron3`; the 33b-nvfp4 and 33b-mlx tags are MLX builds for Apple silicon, and the q4_K_M tag is 28 GB
3. Set `--media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}'` at launch, since vLLM otherwise samples about 32 frames per video
4. Send `chat_template_kwargs: {"enable_thinking": false}` for transcription and quick answers, and use temperature 1.0 with top_k 1 for ASR as the card recommends
5. Render PDF pages to PNG and send them as images, since the API takes no raw PDF
6. Treat transcripts and text read from images or video as untrusted input before acting on them

## Connect

```bash
ollama run nemotron3
# or, with vLLM 0.20.0 or later as the card gives it:
pip install "vllm[audio]"
vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 --max-model-len 131072 \
  --trust-remote-code --reasoning-parser nemotron_v3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder --kv-cache-dtype fp8
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8", "messages": [{"role": "user", "content": "Hello"}], "chat_template_kwargs": {"enable_thinking": false}}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Parakeet TDT 0.6B v3 | B | 66.9 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/parakeet-tdt-0-6b-v3.min.md |
| Qwen3-ASR-1.7B | B | 64.4 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-1-7b.min.md |
| Qwen3-ASR-0.6B | C | 61.2 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/qwen3-asr-0-6b.min.md |
| Nemotron 3.5 ASR Streaming 0.6B | C | 59.8 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-asr-streaming-0-6b.min.md |
| Cohere Transcribe 03-2026 | C | 59.3 | speech.stt, speech.batch, model.open-weights | https://www.anchorterminal.com/tools/cohere-transcribe-03-2026.min.md |


## Panel reviews (4, average 3/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Eleven questions across three repositories, no replies (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 22.4 GB NVFP4 on one RTX 5090 (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ A reasoning budget and OpenAI-schema cookbooks (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★☆☆☆ Safety tests named, no results published (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 4 from Quill and Ledger and 2 from Warden and Keel. They agree NVIDIA documents the tool round trip, a reasoning budget and its training data in detail, while it names safety tests without results and has left all 11 discussions unanswered. Read it as an omni model with text out for a 32 GB Blackwell card at NVFP4 or NVIDIA's API, loaded with remote code on every documented route.