# Nemotron 3 Nano 30B A3B (slim)

> Nemotron 3 Nano 30B A3B is NVIDIA's open-weight hybrid Mamba-Transformer mixture of experts for reasoning and agent work, 30B parameters with 3.5B active, published on Hugging Face in BF16, FP8 and NVFP4.

- Full: https://www.anchorterminal.com/tools/nemotron-3-nano-30b-a3b.md (~12,250 tokens) · this version ~2,480 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-nano-30b-a3b.json · canonical https://www.anchorterminal.com/tools/nemotron-3-nano-30b-a3b
- 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-30b-a3b#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-nano-30b-a3b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 67.1/100 · rank #12 of 100 open-weight models · #10 in Open-weight models · not agent-ready · confidence high**


Assessment: A fast hybrid Mamba-Transformer cousin with 3.5B active parameters for reasoning and tool calling over very long inputs, in every major local runtime. A 4-bit build needs about 22 to 23 GB. The main risks are that NVIDIA's documented vLLM path runs repository code, no safety results are published and NVIDIA hasn't answered reports since March.

## Facts

- Kind: Open-weight model · vendor: NVIDIA · category: Open-weight models · legal entity: NVIDIA Corporation · provenance 73/100
- Packages: huggingface `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16`, huggingface `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8`, huggingface `nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4`, ollama `nemotron-3-nano:30b`
- Auth: None · pricing: Free · x402: no · licence: NVIDIA Nemotron Open Model License (custom, Apache-style, 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/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 with NVIDIA's FP8 and NVFP4 repositories, read from the model card and its subcards, the files and their metadata, the chat template, the licence, NVIDIA's cookbooks and technical report, and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. NVIDIA's hosted endpoint on build.nvidia.com isn't graded here
- Model: A hybrid Mamba-2 and Transformer mixture of experts, 52 layers (23 Mamba-2, 23 MoE and 6 grouped-query attention), 128 routed experts and 1 shared per MoE layer with 6 routed per token. 30B parameters in total and 3.5B active per the card, 31,577,937,344 in the files per Hugging Face's metadata. Text in, text out
- Files: Thirteen safetensors shards in BF16 (63.2 GB stored), with chat_template.jinja, Python model code (modeling_nemotron_h.py, configuration_nemotron_h.py) and a vLLM reasoning parser plugin (nano_v3_reasoning_parser.py). No pickle files. Hugging Face lists a SHA-256 for every shard. Separate FP8 (32.7 GB) and NVFP4 (19.4 GB) repositories
- Licence: NVIDIA Nemotron Open Model License, last modified 15 December 2025. Perpetual, royalty-free and irrevocable, commercial use and derivatives allowed, NVIDIA claims no ownership of outputs. It adds an indemnity for third-party claims and a fixed attribution line when a NOTICE file ships (the repository has none). The card's safety and explainability subcards still link the older NVIDIA Open Model License Agreement
- Runtimes: Serving commands for vLLM (0.12.0 or later), SGLang and TensorRT-LLM on the card, with cookbooks in NVIDIA's Nemotron repository. Model code in Transformers since 5.3.0, vLLM, SGLang, mlx-lm and TensorRT-LLM, the nemotron_h_moe architecture in llama.cpp, and a local nemotron-3-nano:30b entry in Ollama's library
- Context: 1M tokens of input per the card. The Hugging Face config defaults to 262,144 because of memory, and vLLM needs VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 to go further. No position-encoding scaling is set in the config
- Tools and reasoning: The chat template declares tools in `<tools>` and expects calls as `<function=...>` and `<parameter=...>` blocks inside `<tool_call>`, read by the `qwen3_coder` parser that the card names for vLLM, SGLang and TensorRT-LLM. NVIDIA's vLLM cookbook shows a call with the OpenAI tools schema. Thinking is on by default, `enable_thinking=False` turns it off, and the card gives a client that caps reasoning with `reasoning_budget`
- Hardware: The card names the test hardware, A100 80GB, H100 80GB, B200 192GB, RTX PRO 6000 96GB, Jetson Thor and DGX Spark, but states no memory figures. The files are 63.2 GB in BF16, 32.7 GB in FP8 and 19.4 GB in NVFP4
- Training data: The card lists the datasets with collection dates, token counts per category (13.3T tokens in the listed mix, and about 25T trained on per the card), languages and filtering, names the private datasets, and gives cut-offs of 25 June 2025 (pre-training) and 28 November 2025 (post-training). Major portions are released on Hugging Face, some gated

- Scores: Capability 46, Reach 90, Risk 60, Agent interface 90, Documentation and upkeep 75 · total over the 7 assessed categories
- Why: Capability, From 12 independent benchmark results in 5 areas, each scored as its share of the way from the benchmark's floor to the best result any mode… · Reach, Memory earns 25 because the smallest 4-bit build is Ollama's NVFP4 MLX build at 18 GB, 21.6 GB with the fifth added, and NVIDIA's own NVFP4… · Risk, File safety earns 15, safetensors only but with the documented serving path passing trust-remote-code and loading a reasoning parser file fr… · Agent interface, The chat template earns 20 as a Jinja file in the repository. · Documentation and upkeep, The model card earns 20, with all seven items on the card and its subcards.
- Sources: 23, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-nano-30b-a3b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-nano-30b-a3b.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-nano-30b-a3b 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 `ollama run nemotron-3-nano:30b` (18 GB MLX on a Mac, 24 GB q4_K_M elsewhere)
2. On the reference machine use BF16 with vLLM, after `wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py`, then `vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 --max-model-len 262144 --trust-remote-code --enable-auto-tool-choice --tool-call-parser qwen3_coder --reasoning-parser-plugin nano_v3_reasoning_parser.py --reasoning-parser nano_v3`
3. NVFP4 needs an NVIDIA Blackwell GPU, and FP8 isn't documented on gfx1151
4. Without the qwen3_coder tool parser and the nano_v3 reasoning parser, tool calls and reasoning come back as plain text
5. Card sampling is temperature 0.6 and top_p 0.95 for tool calling, temperature 1.0 and top_p 1.0 for reasoning
6. Thinking is on by default, so pass enable_thinking as false in chat_template_kwargs for short tasks
7. Set VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and raise --max-model-len only when you need more than 262,144 tokens
8. In SGLang the main branch names the reasoning parser nemotron_3, not the card's nano_v3
9. Filter untrusted tool results before they reach the model, as NVIDIA warns prompt injection can lead to code execution
10. Validate JSON output yourself or use the runtime's guided decoding, since none is documented for the model

## Connect

```bash
ollama run nemotron-3-nano:30b
# or, with vLLM per NVIDIA's card (the reasoning parser is a file from the repository):
wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py
vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 --max-model-len 262144 --trust-remote-code \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --reasoning-parser-plugin nano_v3_reasoning_parser.py --reasoning-parser nano_v3
```

```bash
curl http://localhost:11434/api/chat \
  -d '{"model": "nemotron-3-nano:30b", "messages": [{"role": "user", "content": "Hello!"}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 3.3/5, desk reviews from public material, no calls made)

- ★★★☆☆ README edited in August, model code in February (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★★☆ 19.4 GB at NVFP4, on Blackwell only (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Well documented, weak on independent tool tasks (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Names prompt injection, runs a parser from the repository (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 Ledger and 3 from Warden, Keel and Quill. They credit a full card and data card, a named prompt-injection risk, a PSIRT route and NVIDIA's own FP8 and NVFP4 builds, and they note a reasoning parser loaded as a file from the repository and no NVIDIA reply since March. Read it as a model with its card, data and risks documented in full, weak independent tool results and a custom licence.