Mistral AI · Open Weights · October 2026

Mistral Large 4: A 1-Trillion-Parameter Model You Can Download — and Probably Should Not Run on Your Laptop

Mistral AI released the weights of Mistral Large 4, code name Le Chonk, on October 29, 2026, after a public preview earlier in October. The headline is real: a flagship model with about 1 trillion total parameters is now open-weight, with download, inspection, customization, and self-hosting on the table. The second headline is the one most guides skip: a 1-trillion-parameter Mixture-of-Experts model is not a download you casually run. This page separates what Mistral actually published from what is still vendor-reported, and gives you a three-question check before you spend anything on hardware.

Last updated: November 1, 2026. Facts checked against Mistral's October 29, 2026 open-weight announcement, the Hugging Face model card for mistralai/Mistral-Large-4-1T-A52B, and Mistral's model documentation. Benchmark figures below are Mistral-reported and marked as such; Mistral says a full technical report is coming soon.

At a Glance

The 3-Question Check Before You Self-Host

Answer these honestly before you download anything. If any answer is no, start on a hosted route and revisit self-hosting only when the answer changes.

What It Is Best At, in Plain Terms

Agentic coding. Mistral positions Large 4 as its best open-weight coding model. The reported numbers are 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, and 48% on the Artificial Analysis Coding Agent Index — effectively tied with the leading open-weight model on that index while using 47% fewer tokens per task on average. Fewer tokens per task is the part that matters for cost: a coding agent that finishes a job in fewer tokens is cheaper to run at scale, even before you compare headline scores.

Defensive cybersecurity. Mistral reports 51% on the Artificial Analysis Cyber Index, which covers discovering vulnerabilities, reproducing and validating them, and patching them without breaking existing functionality, and a 96.6% average refusal rate on malicious cyber requests. Note the distinction Mistral itself draws: the open weights are for defensive work, while a specialized variant for advanced workflows such as penetration testing is offered separately to vetted enterprise design partners, restricted to owned or explicitly authorized systems. Do not read the open release as an unrestricted offensive-security tool.

Long, agentic work. A 1 million-token context window, native tool calling, and three reasoning modes are the practical combination here: the model can hold a large codebase or a long investigation in context, call tools, and switch from a cheap low-effort reply to a high-effort run when the task demands it. That is also where costs hide — a high-effort run over a 1M context is not a cheap run, on the API or on your own GPUs.

How to Deploy It

vLLM (recommended by Mistral). Use the official vLLM Docker image with support for Mistral Large 4. The model card's serving example, with 8-way tensor parallelism, is:

docker pull 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

The two parser flags are required for tool use and for separating reasoning from visible content. If you run short of GPU memory, the model card's advice is to lower --max-model-len and/or --max-num-seqs (concurrency) rather than guessing a smaller checkpoint. A quantized NVFP4 checkpoint is also available with a smaller memory footprint, targeting Blackwell GPUs, and supporting vLLM and SGLang.

SGLang, transformers, and llama.cpp. SGLang is Mistral's other production-ready serving route, with its own cookbook. Transformers and llama.cpp are also listed on the model card for inference and experimentation. For most teams the honest order is: try the API first, then vLLM or SGLang on a server you already run, and only then consider a custom serving stack.

Fine-tuning. The model card lists Axolotl, Unsloth, and TRL as fine-tuning routes, and Mistral Forge for tailoring the model to proprietary data, policies, and detection logic. Fine-tuning a 1T MoE is a different budget from fine-tuning a small model — start with prompting and retrieval on the hosted model before you commit to a training run.

Cost, Honestly

What Is Still Unconfirmed

Three things to keep straight. First, the benchmark figures on this page are Mistral's, and Mistral says a full technical report is coming soon — until that report and independent evaluations land, treat the scores as claims to test, not as settled rankings. Second, third-party hardware figures for running a 1T model vary widely and are not in Mistral's announcement or model card, so this page does not state a minimum GPU count or a RAM figure as fact; the only hardware guidance it relies on is the official 8-way tensor-parallel example and the model card's memory-conservation advice. Third, the specialized cybersecurity variant for vetted enterprise partners is not the same thing as the open weights — do not assume you get that variant by downloading the public checkpoint.

Official sources: Mistral AI — Mistral Large 4 (Le Chonk) is now open weight (October 29, 2026), Hugging Face — mistralai/Mistral-Large-4-1T-A52B model card, and Mistral Docs — Mistral Large 4.

FAQ