Mistral Large 4: API in preview now, open weights due by end of October

On 6 October 2026 Mistral AI launched Mistral Large 4, its largest model, in public preview. The API is available now on Mistral Studio at half price for two weeks; the weights, meaning the model's files, are due by the end of October under a licence that has not yet been published. If you are considering an AI assistant for your business, you can run a measured trial now on non-sensitive data, but do not build a process on it until the licence, the contract and the final version are known.

What Mistral announced

• Mistral Large 4, also called Le Chonk, is a mixture-of-experts model: only part of its parameters work on each request. It has about a trillion parameters (1.05 trillion on the model card), with 49 billion active for each token, or 52 billion counting embeddings and output layers, according to Mistral's Hugging Face page.
• It accepts text and images as input and produces text only, as VentureBeat confirms; the model card lists a context window of 1 million tokens.
• Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres, where the preview also runs, and that the training data covers more than 160 languages, including every official EU language and so Italian.

Claimed and measured performance

The comparisons Mistral publishes are the company's own claims: it says the model is competitive with the strongest open-source models in the world, clearly ahead of any open-weight model developed in the United States or Europe, and state of the art among open models on cybersecurity, finance and law.

Among third-party measurements, Artificial Analysis has published a page for the preview: 38 points on its general index, above the median of comparable models (26). It calls the model very verbose: 200 million output tokens in its test, against a median of 81 million. It also lists a context window of about 524,000 tokens, against the 1 million Mistral states. VentureBeat notes that the 62% on DeepSWE, a coding test, is competitive but does not show a clear lead.

Prices and access

Mistral's changelog gives a 50% discount for two weeks without an end date: if they start on 6 October, they run out around the 20th. A thousand requests of 2,000 input tokens and 500 output tokens cost about $4.8 at list price and about $2.4 at the launch price [illustrative estimate, to be confirmed against your own project].

Data and EU residency

Mistral's help article, updated on 12 August (before the launch), says data is hosted in the EU by default and the US endpoint has to be chosen explicitly. Depending on the feature, however, data may be transferred temporarily outside the EU, to the locations listed in Mistral's Trust Center, under contractual safeguards in line with Article 46 of the GDPR.

Regional inference is an optional service: it limits processing to data centres in the EU and EFTA through the endpoint api.eu.mistral.ai, at a 10% surcharge. Without a region the request goes to the global endpoint, for which Mistral makes no commitment on where processing happens. The page does not name Large 4 and warns that the available models differ from region to region, so check the list on the European endpoint before sending real requests; Agents, Batch and the Files API are not available there.

Mistral says the model will also be available in a European deployment that it runs itself, independent of other digital service providers and under European law.

What it means for a small business

Open-weight means the weights can be downloaded; the licence decides what you can do with them, and Artificial Analysis classes open models with limited commercial use separately. The licence for Large 4 is not yet public: the model card names none, and VentureBeat writes that it will be a custom Mistral licence. In our view hosting the model in-house is out of reach for almost every small business; the advantage of open weights is being able to hand them to different providers, if the licence allows it.

The comparison with the model you use now should be made on the cost of a whole task, on the same examples: output costs more than input, so a verbose model weighs more.

Where we start

In the AI automations we build, the model is chosen after the process has been mapped, on accuracy, cost, speed and operational risk. With Large 4 we would start from two questions: what a single task costs, and which data leave the company, towards which endpoint and under which contract.

Two of our open-source libraries (Apache 2.0) do not depend on the model you pick. Kmemo is a semantic cache for model calls: it recognises a request that has already been answered even when it is reworded, and rejects cases where the similarity is high but the right answer differs, so that, when requests repeat, fewer calls are made. The model goes into the cache's scope, so changing it does not serve one model's answers to another. Requests and answers stay in a store you choose (Redis, PostgreSQL or a file), which belongs on your data map. Kdrant is the Kotlin client for the Qdrant vector database: for an assistant that answers from company documents we propose keeping the index on a Qdrant under your control, so that only the retrieved passages reach the model.

We do not assess the licence in place of a lawyer: when Mistral publishes it, we tell you what changes technically, and the legal reading stays with your adviser.

What we do not know yet

• The licence terms are not public, so commercial use, hosting by third parties and modification remain to be checked.
• The date of the weights is not fixed: end of October for Mistral, the 27th for VentureBeat, the 31st as an estimate on the Hugging Face page.
• The final version may differ from the one you try today, because the training run behind the preview continues.
• The sources report no measurements of quality in Italian and give two different context windows: 1 million tokens for Mistral, about 524,000 for Artificial Analysis.

If you want us to look at your workflow and measure cost per task and where the data goes before choosing the model, the service is described on the page for AI automation and chatbots and you can write to us.

Sources

• Mistral AI: Introducing Mistral Large 4 (6 October 2026)
• Mistral AI Docs: Mistral Large 4 (6 October 2026)
• Mistral AI Docs: Changelog
• Mistral AI Docs: Model lifecycle policy
• Mistral AI Docs: Regional inference
• Mistral AI Help Center: Where do you store my data or my Organization's data?
• Hugging Face: mistralai/Mistral-Large-4.0-1T05-A52B (upcoming release)
• VentureBeat: Mistral debuts Large 4 'Le Chonk', a 1-trillion parameter text output model with high benchmarks planned for open weights release (6 October 2026)
• Artificial Analysis: Mistral Large 4 Preview Intelligence, Performance & Price Analysis