Mistral Large 4: Specs, License & How to Run

Mistral Large 4 (also Le Chonk) is Mistral AI’s ungated natively multimodal Mixture-of-Experts checkpoint. Official hosts we opened on 2026-10-30: the ungated FP8 card mistralai/Mistral-Large-4-1T-A52B (created 2026-09-30T18:35:13Z; last modified 2026-10-29T14:51:07Z; 1,301 likes / 1,239 downloads), the ungated NVFP4 sibling mistralai/Mistral-Large-4-1T-A52B-NVFP4 (created 2026-10-15T19:09:41Z; 50 likes / 315 downloads), the Mistral Ecosystem Open-Weight LICENSE, the OSS post Mistral Large 4 (Le Chonk) is now open weight, and Hugging Face’s Welcome Le Chonk (published 29 October 2026). API gated was false on both cards. Hugging Face pipeline_tag is image-text-to-text.

This page hangs off ChatGPT alternatives, closed LLM APIs, and Claude Code alternatives. Those mappings are editorial (chat / API / coding-agent search). We have no sourced quality comparison against ChatGPT, Claude Code, or Muse Spark Open. Head-to-head facts: Mistral Large 4 vs Muse Spark Open.

The card calls it “our largest and most capable model to date” and a 1-trillion-parameter natively multimodal MoE with 49 billion active parameters per token (52 billion including embeddings and output layers), trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral datacenters in Europe. Those adjectives and the GPU count stay vendor. The Hugging Face blog states two official variants: FP8 and NVFP4.

Do not flatten this card into a consumer GPU story. runsLocally: true means the weights are downloadable without a gate.

What we verified (2026-10-30)

Field	Value	Source
Host (FP8)	huggingface.co/mistralai/Mistral-Large-4-1T-A52B (not gated)	HF API
Created / last modified	2026-09-30T18:35:13Z / 2026-10-29T14:51:07Z	same
Likes / downloads	1,301 / 1,239	same
License tag	other	HF cardData
LICENSE file	Mistral Ecosystem Open-Weight License (modified MIT)	LICENSE
Pipeline tag	image-text-to-text	HF API
Architecture	mistral_large4 / MistralLarge4ForConditionalGeneration	config
Params (card)	~1T total; 49B active/token; 52B including embeddings	card
Quantization (lead card)	FP8 (quant_method: fp8)	config
usedStorage (FP8)	1,060,608,252,777 bytes (~988 GiB / ~1.06 TB)	HF API
NVFP4 sibling	ungated; 50 likes / 315 downloads; usedStorage 610,096,072,460	HF API
Context	1M tokens (card)	card
DSpark	embedded draft head; num_speculative_tokens 3	card
Languages on cardData	38 ISO codes (EN, FR, DE, ES, …; includes EU official set on the card)	HF cardData
Card “160+ languages”	vendor sentence	card

We keep Mistral’s ~1T / 49B / 52B quotes. We do not invent a 1.05T parameter reconciliation. Disk size (~1.06 TB FP8 usedStorage) is not a parameter count.

How to run (from the official card, not tested)

Copied names only. forkalt has not loaded a checkpoint.

Recommended sampling on the card: temperature=1.0, top_p=0.95. Reasoning effort low / medium / high (default high).

vLLM Docker recipe from the card (vllm/vllm-openai:mistral-large-4):

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 \
  --tensor-parallel-size 8 \
  --tool-call-parser mistral \
  --reasoning-parser mistral \
  --enable-auto-tool-choice \
  --gpu-memory-utilization 0.85

To enable the embedded DSpark head, the card adds:

--speculative-config '{"method": "dspark", "num_speculative_tokens": 3}'

The card’s vendor line for DSpark is “up to 1.9x higher throughput.” That stays vendor. Actual gains depend on hardware and workload.

SGLang Docker image named on the card: lmsysorg/sglang:mistral-large-4, with --speculative-algorithm DSPARK and example --tp-size 8. Transformers path on the card uses AutoModelForMultimodalLM from main and torchrun --nproc-per-node 8. The Hugging Face blog also names llama.cpp GGUFs from Unsloth and the llama.cpp team; we did not open those sibling cards in this check.

The card also names Mistral Vibe (mistral-vibe) as a CLI coding assistant that can point at the Mistral API (mistral-large-4) or a local vLLM server. That is a harness label, not a forkalt test, and not a Claude Code clone.

Hosted Mistral API mistral-large-4 is a different artifact from the weight grant.

License

LICENSE is titled Mistral Ecosystem Open-Weight License. Opening sentence: this license is a modified version of the MIT license. Attribution notice: 2026 - Mistral. Hugging Face cardData.license is the opaque label other. Same file is listed on the NVFP4 sibling.

Paraphrase only, not legal advice:

• Include the copyright notice and this permission notice with copies or substantial portions.
• If you or affiliates operate a Model as a Service business, you must enter a separate commercial license with Mistral before using the Model or derivatives for any commercial purpose (contact form on mistral.ai).
• That MaaS requirement does not apply to internal use that does not make the Model available to third parties, or to use through Mistral’s official products.
• “Model as a Service” means giving a third party access for inference or post-training (for instance via API or hosted endpoints) with meaningful control over inputs, parameters, or training data. It does not include end-user products with model capabilities solely embedded in specific features or harnesses, or mere relaying of requests to models hosted by others.

What we did not invent

• VRAM, tokens/s, “frontier / SOTA / best open-weight” stay vendor.
• Card and OSS-blog bench rows stay self-reported (Artificial Analysis indexes, FrontierCode 2.0, Terminal-Bench 4.0, Cyber Index, Surge Code preference, 96.6% malicious-cyber refusal).
• We do not rank Mistral Large 4 against ChatGPT, Claude Code, GLM-5.3, Kimi K3, DeepSeek-V4.1-Flash, or Muse Spark Open.

Related on this site: Muse Spark Open — Mistral Large 4 vs Muse Spark Open — GLM-5.3-Flash — DeepSeek-V4.1-Flash — Kimi K3 — Reflection Beam. Serving: vLLM. UIs: Open WebUI, LibreChat, Jan. Catalog: home. Editorial: About.