# Mistral Large 4 (slim)

> Mistral Large 4 ("Le Chonk") is Mistral AI's open-weight mixture-of-experts model of about 1T parameters with 52B active, taking text and images, with three reasoning efforts and a 1M-token context, published in FP8 and NVFP4.

- Full: https://www.anchorterminal.com/tools/mistral-large-4.md (~10,500 tokens) · this version ~2,380 tokens · JSON https://www.anchorterminal.com/tools/mistral-large-4.json · canonical https://www.anchorterminal.com/tools/mistral-large-4
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-30 · pricing https://www.anchorterminal.com/tools/mistral-large-4#pricing · reviews https://www.anchorterminal.com/tools/mistral-large-4#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 54.6/100 · rank #92 of 117 open-weight models · #72 in Open-weight models · not agent-ready · confidence medium** · same model listed separately: Mistral AI API, Model API (https://www.anchorterminal.com/tools/mistral-api.min.md)


Assessment: A frontier-sized open cousin for coding agents and defensive security, with strict JSON Schema output, parallel tool calls and three reasoning efforts. It needs a multi-GPU server, about 732 GB for the smallest 4-bit build, and vLLM and SGLang run it only from special images. The main risk is a licence that bars Model-as-a-Service businesses without a separate deal.

## Facts

- Kind: Open-weight model · vendor: Mistral AI · category: Open-weight models · legal entity: Mistral AI (RCS Paris 952 418 325) · provenance 84/100
- Packages: huggingface `mistralai/Mistral-Large-4-1T-A52B`, huggingface `mistralai/Mistral-Large-4-1T-A52B-NVFP4`
- Auth: None · pricing: Free · x402: no · licence: Mistral Ecosystem Open-Weight License (modified MIT)
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/mistralai/Mistral-Large-4-1T-A52B (FP8) and Mistral-Large-4-1T-A52B-NVFP4, read from the model cards, the files and their metadata, the chat template, the licence, Mistral's announcements and docs, and the runtimes' own code on 30 October 2026. Nothing was downloaded or run. The hosted model is the mistral-api listing
- Model: A granular mixture of experts, 72 layers with 128 routed experts (4 active per token) and 1 shared, a dense first layer, multi-head latent attention with a sparse attention indexer, and a Pixtral-style vision encoder of 1.6B. About 1T parameters in all (1.05T in Mistral's docs) and 49B active per token, 52B with embeddings and output layers. Text and images in, text out. An embedded DSpark draft head of 3 dense layers proposes 3 tokens at a time for speculative decoding
- Files: 54 safetensors shards in block-scaled FP8 (1,060.6 GB) and 32 in the NVFP4 repository (610.0 GB, NVFP4 experts with FP8 linears), each with chat_template.jinja, params.json, tekken.json and tokenizer.json. No pickle and no Python files. Hugging Face lists a SHA-256 for every shard. No BF16 repository
- Licence: Mistral Ecosystem Open-Weight License, a modified MIT licence. Use, modification, distribution and commercial use are free, except that a licensee whose group operates a Model-as-a-Service business (giving third parties inference or post-training access with meaningful control over inputs, parameters or training data) must take a separate commercial licence before any commercial use. Internal use and use through Mistral's own products are exempt. The card adds a line forbidding use that violates third parties' rights
- Runtimes: Model code in released Transformers (mistral_large4 in v5.22.0) and llama.cpp (LLM_ARCH_MISTRAL_LARGE4 at release b12017), with GGUFs from ggml-org. vLLM and SGLang serve it only from Docker images tagged mistral-large-4; no release of either has the model. Ollama lists it as a cloud model only, and LM Studio has no entry
- Context: 1,048,576 tokens in config.json, reached with YaRN (factor 2048 over an original 512 positions) and NoPE on every fourth layer. Mistral says the checkpoint supports a 1M window and doesn't say what length it was trained at
- Tools and reasoning: The template declares tools in [AVAILABLE_TOOLS] and renders calls as [TOOL_CALLS]name[ARGS]arguments, several per turn, with results in [TOOL_RESULTS]. The card names vLLM's mistral tool and reasoning parsers and SGLang's auto parsers, and shows OpenAI-style tool calls and strict JSON Schema output through response_format, which the template passes to the model in [MODEL_SETTINGS]. reasoning_effort is low, medium or high (default) per request and can't be turned off; the reasoning comes back in message.reasoning
- Hardware: The card's commands use 8 GPUs with tensor parallelism and says NVFP4 targets Blackwell GPUs. vLLM's recipe validates GB300, B300, B200 and H200 (FP8 at about 128 GiB of weights per GPU on 8 GPUs, NVFP4 about 76 GiB) and AMD MI350X. Nothing fits a 128 GB machine; the smallest 4-bit build needs about 732 GB
- Recommended settings: temperature 1.0 and top_p 0.95 at every effort, high effort for agentic use and complex prompts, low for quick replies, a system prompt that sets the environment, and as few tools as the task needs, per the card
- Training data: Not described on the card. Mistral's preview post says it was trained from scratch on 3,800 Grace Blackwell GPUs in Mistral's own European datacentres, that a significant share of the data spans more than 160 languages, and that post-training used SFT and reinforcement learning in the same environments Mistral Forge sells to customers. No sources, sizes or cut-off. A technical report is described as coming soon

- Scores: Capability 32, Reach 50, Risk 75, Agent interface 95, Documentation and upkeep 64 · total over the 7 assessed categories
- Why: Capability, From 5 benchmark results (1 submitted by the publisher, at half weight) in 2 areas, each scored as its share of the way from the benchmark's… · Reach, Memory scores 0, since the smallest 4-bit build, Mistral's NVFP4, is 610 GB, about 732 GB with a fifth added, far above 256 GB. · Risk, File safety scores 25, safetensors only and loaded by Transformers' own code with a Jinja template. · Agent interface, The chat template is in the repository (20). · Documentation and upkeep, The card covers intended use, architecture, parameters, context and languages, and misses limitations and a cut-off (14).
- Sources: 25, open questions: 5, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/mistral-large-4.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/mistral-large-4.svg` or a link to https://www.anchorterminal.com/tools/mistral-large-4 from a page on mistral.ai 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.
- Support: Discord https://discord.gg/mistralai; https://help.mistral.ai/en/articles/347458-how-do-i-contact-support; https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/discussions; more in the full twin

## Before you call it

1. Serve the NVFP4 build on Blackwell with `vllm/vllm-openai:mistral-large-4 --model mistralai/Mistral-Large-4-1T-A52B-NVFP4 --tensor-parallel-size 8 --kv-cache-dtype fp8_ds_mla --tool-call-parser mistral --reasoning-parser mistral --enable-auto-tool-choice`; released vLLM doesn't load it
2. On H200 use the FP8 build with tensor parallelism of 16 or 8 by 2 pipelines, per vLLM's recipe
3. Pass `reasoning_effort` low, medium or high (default high); any other value makes the template raise an error, and there is no off switch
4. Read the reasoning from `message.reasoning`; `max_completion_tokens` counts reasoning tokens too
5. Use temperature 1.0 and top_p 0.95 at every effort, per the card
6. Ask for JSON with `response_format` type json_schema and strict true; the template passes the schema to the model
7. The template rejects a tool message straight after a user message and a conversation that doesn't start with a user or system message
8. Add `--speculative-config '{"method": "dspark", "num_speculative_tokens": 3}'` to use the embedded draft head
9. If you'll offer it to others as a hosted model, read the licence first; Model-as-a-Service use needs Mistral's commercial licence

## Connect

```bash
# NVFP4 on 4 or 8 Blackwell GPUs, with the card's vLLM image (no vLLM release supports it yet):
docker run --runtime nvidia --gpus all -v ~/.cache/huggingface:/root/.cache/huggingface -p 8000:8000 --ipc=host \
  vllm/vllm-openai:mistral-large-4 --model mistralai/Mistral-Large-4-1T-A52B-NVFP4 \
  --tensor-parallel-size 8 --kv-cache-dtype fp8_ds_mla --tool-call-parser mistral \
  --reasoning-parser mistral --enable-auto-tool-choice --gpu-memory-utilization 0.85
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "mistralai/Mistral-Large-4-1T-A52B-NVFP4", "reasoning_effort": "medium", "temperature": 1.0, "top_p": 0.95,
       "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 (0, desk reviews from public material, no calls made)