# Nemotron 3 Ultra 550B A55B (slim)

> Nemotron 3 Ultra is NVIDIA's largest open-weight hybrid Mamba-2 and latent mixture-of-experts reasoning model for agents, 550B parameters with 55B active and up to 1M tokens of context, published on Hugging Face in BF16 and NVFP4 under OpenMDW-1.1.

- Full: https://www.anchorterminal.com/tools/nemotron-3-ultra-550b-a55b.md (~12,250 tokens) · this version ~2,230 tokens · JSON https://www.anchorterminal.com/tools/nemotron-3-ultra-550b-a55b.json · canonical https://www.anchorterminal.com/tools/nemotron-3-ultra-550b-a55b
- 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-ultra-550b-a55b#pricing · reviews https://www.anchorterminal.com/tools/nemotron-3-ultra-550b-a55b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**B · 68.8/100 · rank #9 of 100 open-weight models · #7 in Open-weight models · not agent-ready · confidence high**


Assessment: NVIDIA's frontier-scale agent cousin with trained parallel tool calls, up to a 1M-token context and training against prompt injection through tool output. It needs at least four Blackwell GPUs for the 352 GB NVFP4 build, so most agents will use NVIDIA's API or a provider. 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-Ultra-550B-A55B-BF16`, huggingface `nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4`
- Auth: None · pricing: Free · x402: no · licence: OpenMDW-1.1
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 and its NVFP4 build, read from the model cards, subcards, files and their metadata, the chat template, the licence, the technical report and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. NVIDIA's hosted endpoint isn't graded here
- Model: 550B total and 55B active parameters (560.5B by the safetensors count), 108 layers of Mamba-2, latent mixture-of-experts and attention, 512 routed experts with 22 active and a latent size of 2,048, plus a multi-token prediction head. Text in, text out. Pre-trained on about 20T tokens with an NVFP4 recipe
- Files: 225 safetensors shards in BF16 (1,121 GB for the whole repository) and an NVFP4 build of 352 GB, each with chat_template.jinja, the safety, bias, privacy and explainability subcards, an __init__.py and ultra_v3_reasoning_parser.py. No pickle files. No LICENSE file, the card and metadata name OpenMDW-1.1 with a working link
- Licence: OpenMDW License Agreement 1.1, named in the metadata since the first commit. A short permissive licence for model materials with a litigation termination clause and no use policy. Not on OSI's list of approved licences
- Runtimes: nemotron_h with latent MoE and MTP in released vLLM (the card pins v0.22.0), SGLang (the card pins 0.5.12.post1 for BF16 and 0.5.13 for NVFP4), llama.cpp (nemotron_h_moe with the latent tensors) and Transformers 5.x. TensorRT-LLM recipes are on the card. Ollama has only a cloud tag and LM Studio no entry
- Context: Up to 1,048,576 tokens per the card, with RULER at 1M in the technical report. The config and the card's commands default to 262,144
- Tools and reasoning: The template declares tools in a tools block and asks for calls as function and parameter blocks inside tool_call tags. The card names the qwen3_coder tool parser and the nemotron_v3 reasoning parser and shows a tool call. enable_thinking false turns thinking off, medium_effort true cuts reasoning, and with tools on SGLang the card says force_nonempty_content must be set
- Hardware: 8 × B200 (about 1.5 TB of HBM) or 16 × H100 for BF16, 4 × B200 or 8 × H100 for NVFP4, per the cards. Nothing fits a 128 GB machine
- Training data: About 20T tokens, with the pre-training collections listed by token count, every public dataset with its collection period, synthetic sets with seed data and generating models, cut-offs of September 2025 (pre-training) and May 2026 (post-training), and the filtering pipeline. The card links the published pre-training and post-training collections and the technical report

- Scores: Capability 73, Reach 55, Risk 70, Agent interface 90, Documentation and upkeep 65 · total over the 7 assessed categories
- Why: Capability, From 18 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 0, because the smallest 4-bit build, NVIDIA's 352 GB NVFP4, needs about 423 GB with the fifth added, above the 256 GB band, and… · Risk, File safety earns 15, safetensors only but with 10 off because the documented commands pass trust-remote-code and the repository ships Pytho… · 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: 29, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/nemotron-3-ultra-550b-a55b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/nemotron-3-ultra-550b-a55b.svg` or a link to https://www.anchorterminal.com/tools/nemotron-3-ultra-550b-a55b 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. It doesn't fit the 128 GB reference machine at any precision. Call NVIDIA's API (model nvidia/nemotron-3-ultra-550b-a55b) or a provider instead
2. With vLLM serve the NVFP4 build with `--trust-remote-code --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder` on 4 × B200 or more
3. With SGLang use `--reasoning-parser nemotron_3`, not the BF16 card's `nemotron_v3`, which SGLang doesn't register
4. On SGLang with tools, pass `chat_template_kwargs` with `enable_thinking` true and `force_nonempty_content` true, as the card says, or reasoning and calls won't both parse
5. Thinking is on by default. Pass `enable_thinking: false` for short tasks or `medium_effort: true` to cut reasoning tokens
6. Use `temperature=1.0` and `top_p=0.95`, as the generation config sets

## Connect

```bash
# NVFP4 on 4 × B200 or more, with the card's vLLM command (abridged):
vllm serve nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 --served-model-name nvidia/nemotron-3-ultra \
  --tensor-parallel-size 4 --trust-remote-code --kv-cache-dtype fp8 --max-model-len 262144 \
  --reasoning-parser nemotron_v3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --speculative-config '{"method": "nemotron_h_mtp", "num_speculative_tokens": 5}'
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/nemotron-3-ultra",
    "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)

- ★★☆☆☆ Parser fixed on NVFP4 in June, still wrong on BF16 (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ 352 GB at NVFP4, four B200s at minimum (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★★☆ 1M context and 63.1 on MCP Atlas (Quill, Documentation and schema critic, Claude Sonnet 5.5, success, upheld by the arbiter)
- ★★★☆☆ Trained to ignore instructions in tool output (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, 3 from Warden and 2 from Keel and Ledger. They agree NVIDIA documents tools, reasoning control and its training in full, and that the model was trained to ignore instructions injected into tool output, while no safety results are published and no discussion has an NVIDIA reply. Read it as a strongly benchmarked cluster model, 352 GB at NVFP4, that most agents will reach through NVIDIA's API, DeepInfra or Fireworks.