When to Download Mistral Large 4

What Mistral shipped on October 29

Mistral published the weights for Mistral Large 4, nicknamed Le Chonk. The public preview opened on October 6. The October 29 checkpoint is a further-trained model, and Mistral's legal page lists the preview as retired. The download is mistralai/Mistral-Large-4-1T-A52B on Hugging Face. Mistral says to serve it with vLLM or SGLang.

The model card describes a natively multimodal mixture-of-experts model: 1.05 trillion total parameters, 52 billion active, a 1.6 billion parameter vision encoder, and a 1 million token context. The API name on the weights post is mistral-large-4-0. Mistral says the API price starting October 29 is $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens.

Mistral Large 4 is in this directory. These figures are Mistral's. They are not a CuratedAI test of the weights, the bill, or a self-hosted run.

The licence is open weights with one commercial condition

Mistral calls the terms the Mistral Ecosystem Open-Weight License, a modified MIT licence. The weights post says you may use, modify, fine-tune, and commercialize the model. A business that operates a model-as-a-service offering must get a separate commercial licence from Mistral before using it that way.

That split is the reason to read the licence before you announce a hosted product. A team running the weights on its own machines, or shipping a product that is not itself a model API, is the case Mistral describes as free. Reselling the model as an API is the case that needs the extra licence. This page is not legal advice. The licence file is the source.

Read the new scores as the weights checkpoint

The October 6 preview and the October 29 weights are different checkpoints. Mistral's preview post reported 61.7 percent on DeepSWE v1.1, 28.3 percent on Terminal-Bench 4.0, and a Coding Agent Index of 49.8 percent. The weights post reports a different set, after three more weeks of reinforcement learning. Do not paste the two sets into one row.

Coding Agent Index. Mistral says the weights checkpoint scores 48 percent, tied with GLM-5.3, while using 47 percent fewer tokens per task on average.
Terminal-Bench 4.0. Mistral says the score rose 13 points from the preview, to 41 percent, and that output tokens fell by 30 percent. It calls that the lead among open-weight models.
FrontierCode 2.0. On Cognition's private benchmark, Mistral says Large 4 leads at 19.7 percent. That test is not public.
Reasoning modes. Low is for simple tasks and replaces a previous "None" mode. Medium is for everyday tasks. High is the default, for work that needs more care. The token savings above depend on which mode you call.

On defensive cyber tests, Mistral says the weights checkpoint scores 51 percent on the Artificial Analysis Cyber Index, and that its average refusal rate on malicious cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is 96.6 percent. A separate, more capable cybersecurity variant is not in the public download. Mistral says that variant is for vetted design partners, under contract, and only on systems they own or were authorized to test. This page does not describe how those tests are run.

Match the job to a download or an API call

The download is for control: where the model runs, and whether a vendor can turn it off. The API is for a job you want to try before you buy a cluster.

You need the model on your own machines

Download the Hugging Face repo and serve it with vLLM or SGLang. Mistral also names an NVIDIA NIM container and Microsoft Foundry for teams that want a managed deployment with a private option. Budget for a trillion-parameter model. Fifty-two billion active parameters does not make the checkpoint small.

You want to try coding or documents first

Call mistral-large-4-0 from Mistral Studio, the API, or Le Chat. Start on Medium for ordinary tasks and move a failed sample to High. Cached input at $0.07 only helps when the same prefix is actually cached.

You would resell the model as an API

Stop and get the commercial licence. The weights post is explicit that model-as-a-service is the exception to the free commercial terms.

You defend software you own

Use the public checkpoint only on systems you are allowed to test, and keep a person on the patch. The specialized cyber variant is a partner program, not a toggle in the public repo. A blog score is not permission to probe someone else's network.

Where this release is easy to overread

"Best open-weight model for coding" is Mistral's summary of its own tables. The Coding Agent Index line is a tie with GLM-5.3 at 48 percent, plus a token-efficiency claim. It is not a claim that Large 4 replaced Claude Opus 5.5 or a closed frontier model on every coding job. Surge's blind code comparison, also on the weights post, still puts Opus 5.5 ahead of Large 4.

The October 6 directory notes that said the weights were still promised describe the preview. The preview model is retired. Prices of $1.36 and $4.18 per million tokens were the earlier quote. The weights post is the rate card to use now, and it can change.

The same day's briefing also covers a personal Claude agent found on a Pentagon network, a CrowdStrike report on DeepSeek used against more than 200 organizations, and Amazon's cloud results. Those are news items. They are not features of this download. Read them in the October 29 briefing.

What to do this week

If you need a European model on hardware you control, download the checkpoint and price a real serving stack before you promise it to anyone. If you only need to see whether it finishes your repo, call mistral-large-4-0 on ten tasks you can check, record finishes and tokens, and keep the model you already trust on anything the sample gets wrong. For the rest of the day, read the October 29 briefing. Reviewed October 30, 2026, from Mistral's weights post, the model card, and the licence file.

Questions

When should you download Mistral Large 4 instead of calling the API?: Mistral published the weights for Mistral Large 4, nicknamed Le Chonk.