Mistral Large 4 entered public preview on 6 October 2026. The European AI company Mistral calls it "our largest and most capable model to date": a model with about a trillion parameters, trained in the company's own European data centres, available now through its API and due to have its weights published by the end of October. For Australian businesses it is not a reason to change tools this week. It is a sign that capable AI models are no longer a choice between a handful of American services and a handful of Chinese ones.
What happened
Mistral published the Mistral Large 4 announcement on 6 October and opened a preview of the model on its developer platform, Mistral Studio. The post describes the model, which the company has nicknamed "le Chonk", as natively multimodal, meaning it was built from the start to handle more than text.
Three statements in the announcement stand out.
- The weights are coming. "We will release the weights by the end of the month," the post says. Weights are the numbers a model learns in training. Publishing them lets other organisations run the model on their own equipment.
- It was built in Europe. Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centres in Europe, and that the preview is served from the same infrastructure.
- It is a preview. Mistral says the model "continues to improve rapidly as we refine it".
The release follows a funding round. In September Mistral announced a 3 billion euro Series D led by Samsung Electronics, which it described as the largest equity round completed by a European technology company. The Large 4 post presents the model as the first milestone on the roadmap that money pays for.
Reporting on the launch for Silicon UK, Matthew Broersma placed it in a year in which Chinese companies have offered open-weight models that narrow the gap with closed models from OpenAI and Anthropic. According to his report, Mistral's pitch is that it is the strongest open-weight model developed outside China.
Key details
| Item | Detail |
|---|---|
| Status | Public preview from 6 October 2026 |
| Size | About 1 trillion parameters (1.05 trillion in Mistral's model documentation), of which 52 billion are active at a time |
| Context window | 1 million tokens, according to Mistral's model documentation (a token is a word or part of a word) |
| List price | US$1.36 per million input tokens and US$4.18 per million output tokens |
| Languages | More than 160, including every official language of the European Union |
| Where it runs | Mistral's European infrastructure now; Mistral says multiple regions worldwide will follow, including a European deployment it operates end to end |
| Weights | Promised by the end of October 2026 |
| Licence | Not stated in the announcement or the model documentation |
Mistral also published test results. They come from the company and the evaluators it chose, so treat them as claims until others repeat them. A few examples:
- In a blind coding evaluation scored by people at Surge AI, the preview rated 3.74 out of 5, second of five models, behind Claude Opus 5 on 4.22 and ahead of GLM-5.3 and Kimi K3.
- On AutomationBench, a set of 657 business workflows, Mistral reports 59.9 per cent.
- On Lakera's B3 attack-resistance benchmark, Mistral reports that the model resisted 93.3 per cent of attacks.
Why it matters
There is a third source of capable models. As Silicon UK describes it, the closed models to beat come from OpenAI and Anthropic, and this year's open-weight challengers have come from Chinese companies. A European company shipping a model at this scale, on its own hardware, gives buyers and software makers another serious option.
Where a model runs is becoming a selling point. Mistral stresses that it operates the European service itself, from data centre to model. That will not decide much for a cafe group in Parramatta. It can matter to an Australian firm with European customers, or to any business whose clients ask where their information is processed and who controls the servers.
Open weights are promised, with the terms still unknown. If the weights arrive as planned, hosting companies and large organisations will be able to run the model in a private cloud or on their own equipment. The announcement does not name a licence, and "open-weight" on its own says little about what is permitted. Another large open-weight release this year, Moonshot AI's Kimi K3, shipped under a custom licence with commercial conditions, so the detail is worth waiting for.
The list price is close to an open-weight rival's. For comparison, DeepSeek's pricing page lists its V4 Pro model at US$1.32 per million input tokens that are not cached and US$3.96 per million output tokens at peak times, and half that off-peak.
It is honest about not being first. By Mistral's own figures the preview sits behind Claude Opus 5 in the blind coding evaluation. The claim is that it is competitive and open, not that it leads.
What this means for businesses
Most small and medium businesses will never call Mistral's API directly. Releases like this reach them through the software they buy and the systems that developers build for them. The useful response is to keep your options open.
- Do not move anything important onto a preview. A preview can change in behaviour, price and limits. Try it on low-risk tasks if you are curious.
- Ask where requests are processed. If data location matters to you or your clients, get the region in writing from whichever supplier you use, and do not rely on a general impression of where a company is based.
- Wait for the licence before planning anything around the weights. The terms decide who may host the model and for what.
- Keep the model swappable. When a developer connects an AI model to your systems, ask for it to be a setting that can be changed and not something buried in the code. Prices and rankings keep moving.
- Test in the languages you use. If you deal with customers or suppliers in other languages, check the quality in those languages yourself. A count of supported languages says little about any one of them.
- Judge on your own tasks. Use a real quote, a real customer email or a real report, and compare the results side by side.
We help businesses plan and connect the systems they run on. Our IT consulting service can help you compare suppliers and their terms, our business systems and integrations work covers how your software connects to outside services, and cloud migration and hosting covers where your systems and data should live. To discuss your situation, send us an enquiry.
Key takeaways
- Mistral Large 4 went into public preview on 6 October 2026, with weights promised by the end of the month.
- It has about a trillion parameters and was trained in Mistral's own European data centres.
- List pricing is US$1.36 per million input tokens and US$4.18 per million output tokens.
- The published results come from Mistral and evaluators it chose, and those figures put it behind Claude Opus 5 on coding quality.
- The licence has not been announced, so wait for it before making plans that depend on the weights.
Frequently asked questions
What is Mistral Large 4?
It is the largest AI model released so far by the European company Mistral. It handles more than text, works in more than 160 languages and is available as a public preview through Mistral's API.
Is Mistral Large 4 open source?
Mistral describes it as open-weight and says the weights will be released by the end of October 2026. The announcement does not state a licence, so whether it meets an open-source definition is not yet known.
How much does Mistral Large 4 cost?
The announcement lists US$1.36 per million input tokens and US$4.18 per million output tokens for the preview on Mistral's API. Prices from other hosts, once the weights are public, may differ.
Where is Mistral Large 4 hosted?
Mistral says the preview is served from its own data centres in Europe, the same infrastructure used to train the model. It says more regions worldwide will follow.
Should a small business switch to Mistral Large 4?
Not on the strength of a preview. It is worth testing on low-risk tasks if location or openness matters to you, and worth making sure your systems can change models easily when the picture is clearer.