Mistral Large 4 is out in preview - a trillion-parameter open model, with the weights due 27 October

Mistral released Mistral Large 4 in public preview today, 6 October. It is a mixture-of-experts model with 1.05 trillion total parameters and 49 billion active per token, according to Mistral's announcement. It takes text and images, returns text, and Mistral says it will publish the weights on 27 October. The company's own nickname for it is "Le Chonk".

For now it is an API you can call through Mistral Studio as `mistral-large-4`. The open-weight part, which is the whole reason this model matters, has not shipped yet.

What is actually new

Two things.

First, size in the open. Most open-weight models a business could self-host have been either small and capable, or large and Chinese. Mistral's pitch is that this is the strongest open-weight model built outside China. VentureBeat's coverage puts it at 62% on the DeepSWE software engineering benchmark, ahead of DeepSeek V4 Pro (57%) and Qwen 3.8 Max (51%). On AutomationBench, which Mistral describes as 657 business workflows across Gmail, Google Sheets, Slack and Salesforce, it reports 59.9%, ahead of Kimi K3 and DeepSeek V4 Pro.

Second, price. The list price on Mistral's announcement is $1.36 per million input tokens and $4.18 per million output. During the preview, the docs show half that: $0.68 input, $2.09 output, $0.07 for cached input. For comparison, Claude Sonnet 5.5 and GPT-6.1 Sol both list at $2 input and $10 output. On output-heavy agent work, Mistral's list price is less than half.

Now the marketing filter. Every benchmark above comes from Mistral. VentureBeat notes the live DeepSWE leaderboard has Claude Opus 5, GPT-6 Astra and Gemini 3.8 Flash around 74%, so "best open model" and "best model" are not the same claim. Mistral says the model has a one-million-token context window; independent testing reported by OrcaRouter found 524,288. And in a blind human evaluation by Surge AI, it scored 3.74 out of 5 on coding quality, second to Claude Opus 5 at 4.22. That is a good result. It is also second.

What it means for a business owner

The useful question is not "is it smarter than Claude". It is "do I need a model I can hold".

There are three kinds of business where that answer is yes. Companies under a contract or regulator that says data must not leave a specific jurisdiction. Companies whose clients ask, in procurement questionnaires, which US providers touch their data. And companies building automation they expect to run for years, who do not want a vendor's next price change or model retirement to break it.

Until now, those businesses had a bad trade. Use a frontier API and accept the dependency, or self-host a smaller open model and accept weaker results on multi-step work. Mistral Large 4 narrows that gap. Mistral also says it runs a European deployment "end-to-end, independently" under European law, which gives EU firms a middle option before the weights even land.

The workflows this changes are the ones where you already trust a model to read, decide and act across tools: triaging a support inbox into a CRM, extracting fields from contracts and invoices, reconciling spreadsheets against an accounting system. Those are what AutomationBench claims to measure. If the 59.9% holds up in independent testing, an open model is now in range for that work.

The honest caveat

Open weights does not mean cheap to run. Only 49 billion parameters are active per token, but all 1.05 trillion still have to sit in memory. Even compressed to 8 bits, that is roughly a terabyte of weights before you serve a single request. This is a multi-GPU server or a rented cluster, not a box under someone's desk. For most small businesses, "we could self-host it" will stay theoretical, and the API is what you will actually use.

The licence is also unnamed. VentureBeat describes it as a custom Mistral licence, and Mistral's announcement does not give terms. Until you have read it, you do not know whether commercial self-hosting is allowed on the terms you need. And Mistral itself concedes the model still trails closed frontier models on coding.

Finally, it is a preview. Pricing is promotional, the context limit is disputed, and the weights are a promise with a date on it.

What to do about it

If data residency or vendor lock-in has ever blocked an automation project for you, take one real workflow, ideally one with a known correct answer you can check against, and run it through the preview API this month while it is half price. Compare the output against whatever model you use today.

Do not commit infrastructure until 27 October passes, the weights are actually downloadable, and you have read the licence. If all three check out, you will have the first credible option to own the model as well as the automation.