# Nemotron 3 Super 120B A12B (slim)

> Nemotron 3 Super is NVIDIA's open-weight hybrid Mamba-2 and mixture-of-experts reasoning model for agents, 120B parameters with 12B active, published on Hugging Face in BF16, FP8 and NVFP4 under the NVIDIA Nemotron Open Model License.

- Full: https://www.anchorterminal.com/tools/nemotron-3-super-120b-a12b.md (~12,450 tokens) · this version ~2,330 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-super-120b-a12b.json · canonical https://www.anchorterminal.com/tools/nemotron-3-super-120b-a12b
- 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-super-120b-a12b#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-super-120b-a12b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 66.2/100 · rank #15 of 100 open-weight models · #12 in Open-weight models · not agent-ready · confidence high**


Assessment: A long-context agent cousin with trained parallel tool calls, a reasoning switch and a full account of its training data, under a free commercial licence. It needs about 100 GB of GPU memory, a 128 GB unified machine or Blackwell hardware for NVIDIA's NVFP4. The main risks are serving commands that run repository code and no published safety results.

## Facts

- Kind: Open-weight model · vendor: NVIDIA · category: Open-weight models · legal entity: NVIDIA Corporation · provenance 73/100
- Packages: huggingface `nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16`, huggingface `nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8`, huggingface `nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-NVFP4`, ollama `nemotron-3-super`
- Auth: None · pricing: Free · x402: no · licence: NVIDIA Nemotron Open Model License
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16, with the FP8 and NVFP4 repositories, read from the model card and its subcards, the files and their metadata, the chat template, the licence, the technical report, NVIDIA's cookbook repository 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: 123,611,033,088 parameters in the BF16 files (the card says 120B, 12B active). 88 layers, 40 Mamba-2, 40 LatentMoE and 8 attention, 512 routed experts and 1 shared, 22 routed per token, routed in a 1,024-wide latent space, plus a multi-token prediction head. Pre-trained in NVFP4. Text in, text out
- Files: 50 safetensors shards in BF16 (247.2 GB), separate FP8 (128.4 GB) and NVFP4 (80.3 GB) repositories, a Base checkpoint and an MTPv2 head in their own repositories. No pickle files. Python model code (modeling_nemotron_h.py) and a vLLM reasoning-parser plugin sit beside the weights. Hugging Face lists a SHA-256 for every shard
- Licence: NVIDIA Nemotron Open Model License, version of 15 December 2025, full text public. Commercial use, derivatives and redistribution free, perpetual and irrevocable. No acceptable use policy. The licence ends for anyone who sues claiming the model or its output infringes, and the user indemnifies NVIDIA against third-party claims
- Runtimes: Serving commands on the card for vLLM 0.18.1, SGLang (a nightly container) and TensorRT-LLM 1.3.0rc8, with NVIDIA's cookbooks. Model code in Transformers since 5.3.0, vLLM, SGLang and mlx-lm, the nemotron_h_moe architecture with Super's latent tensors in llama.cpp, and a local nemotron-3-super entry in Ollama's library
- Context: Up to 1,048,576 tokens per the card, with RULER results at 1M. The configuration defaults to 262,144 to save memory, and the card gives the flags that raise it
- 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 in vLLM, SGLang and TensorRT-LLM. Thinking is on by default, `enable_thinking: false` turns it off, `low_effort: true` shortens it, and the card gives a client for a hard `reasoning_budget`. Recommended sampling `temperature=1.0`, `top_p=0.95` for every task
- Hardware: BF16 needs at least 8× H100-80GB, or 2× B200 or B300, per the card. The NVFP4 build runs on 1× B200 or 1× DGX Spark per its card. Tested on 1 to 8 H100 or H200 and on GB200. Fits in 128 GB of memory with the NVFP4 build or Ollama's 87 GB build
- Training data: A data card in the model card. Pre-training cut-off June 2025 and post-training February 2026, 153 datasets listed by name with token counts for the main collections, the synthetic-data teacher models named, and filtering steps described. Major parts of the corpora are released on Hugging Face

- Scores: Capability 54, Reach 75, Risk 65, Agent interface 90, Documentation and upkeep 65 · total over the 7 assessed categories
- Why: Capability, From 10 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 10, because the smallest 4-bit build, NVIDIA's 80.3 GB NVFP4, needs about 96 GB with the fifth added, which fits the reference… · Risk, File safety earns 15, safetensors only but with 10 off because the documented serving commands pass trust-remote-code and the repository shi… · Agent interface, The chat template earns 20, a Jinja file in the repository. · Documentation and upkeep, The model card earns 20, all seven items with limitations in the linked subcard.
- Sources: 24, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-super-120b-a12b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-super-120b-a12b.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-super-120b-a12b 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. On the reference machine start it with `ollama run nemotron-3-super`, the 87 GB q4_K_M build, the only one that fits (about 104 GB with the margin)
2. Raise the Strix Halo GPU memory limit first with `amd-ttm --set 100`, since the default of about 62.8 GiB is far below what it needs
3. With vLLM serve it with `--reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder` (SGLang uses `--reasoning-parser nemotron_3`)
4. Use `temperature=1.0` and `top_p=0.95` for every task, as the card says
5. Thinking is on by default. Pass `enable_thinking: false` or `low_effort: true` in `chat_template_kwargs` for short tasks
6. For coding agents pass `force_nonempty_content: true` in `chat_template_kwargs`, so a cut-off reasoning trace still returns content
7. Raise context past 262,144 only with the card's flags (`VLLM_ALLOW_LONG_MAX_MODEL_LEN=1`, `--max-model-len 1048576`)

## Connect

```bash
ollama run nemotron-3-super
# or, with the card's vLLM command (abridged):
pip install vllm==0.18.1
vllm serve nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 --served-model-name nvidia/nemotron-3-super \
  --tensor-parallel-size 8 --max-model-len 262144 --kv-cache-dtype fp8 --mamba-ssm-cache-dtype float32 \
  --trust-remote-code --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/nemotron-3-super",
    "messages": [{"role": "user", "content": "What is the capital of Japan?"}],
    "temperature": 1.0, "top_p": 0.95,
    "chat_template_kwargs": {"enable_thinking": false}
  }'
```

## 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 2.8/5, desk reviews from public material, no calls made)

- ★★☆☆☆ GA in March, its own parser pull request open since May (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 80.3 GB at NVFP4, 247.2 GB at BF16 (Ledger, Cost analyst, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Two parser instructions on one card (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ PSIRT on the card, model code in 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 3 from Warden, Quill and Ledger and 2 from Keel. They credit a vulnerability route on the card, a full card and data card, reasoning off, low effort or budgeted, and NVIDIA's own NVFP4 and FP8 builds, and they find the parser story muddled, from a plugin in the repository to a removed deployment guide. Read it as a well-documented agent model for a 128 GB machine or Blackwell hardware, with no NVIDIA replies in its discussions.