Mistral AI released a public preview of Mistral Large 4 on October 6, 2026. It is the company's largest model, with 1 trillion parameters. A preview API is open today on Mistral Studio, and Mistral says the weights will follow by the end of the month. It matters to any business weighing which AI model to build on.
What we know
- What it is: a natively multimodal model with 1 trillion parameters, 49 billion of them active, according to Mistral's announcement. The model card puts the total at 1.05 trillion, plus a 1.6-billion-parameter vision encoder.
- Names: Mistral shortens it to ML4 and nicknames it "le Chonk".
- What you can use today: the preview API, through Mistral Studio, under the model ID "mistral-large-4", per the model card.
- API price: $1.36 per million input tokens and $4.18 per million output tokens, per the announcement and the model card.
- Context window: 1 million tokens, per the model card.
- The weights: not out yet. Mistral says it will release them by the end of October, with more detail on the architecture and more benchmarks.
- In the meantime: Mistral says it is red-teaming the model with cybersecurity companies, vetted partners and state authorities, who get the same model with reduced moderation and expanded cyber capabilities.
- Where it was trained: from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers in Europe. The preview is served from that same infrastructure, the announcement says.
- Languages: a significant share of the training data was multilingual, covering more than 160 languages, including every official language of the European Union.
- Benchmarks: they are the ones Mistral publishes. It credits some to third parties: Artificial Analysis, vals.ai and Surge AI. It says the model ranks among the top five worldwide on the Artificial Analysis Cyber Index, and placed second of five models in a blind human evaluation of code quality, at 3.74 out of 5.
- What isn't known yet: which license the weights will carry, or what hardware it takes to run them. The announcement doesn't say, and the model card lists the license as "Coming soon" and gives no figure for the GPU memory needed.
What changes and what doesn't
The menu changes. There is a new very large model, trained in Europe, that you can test by API today and whose weights are due for download. For a business, that is one more option to compare on price and quality.
"Open weights" does not mean "runs on any server". A 1-trillion-parameter model will not fit on ordinary web hosting or an office computer. For a small company, the realistic route is still the API or a provider that hosts the model. That is an observation from this article: Mistral has not published hardware requirements.
It is not a finished model either. It is a preview. Mistral says training is still running and that it expects improvements in the coming weeks. The performance figures are the announcement's own, and independent tests of the final version are still missing.
On cybersecurity, Mistral argues that several closed models score near zero on one test because they refuse the task, and it names Claude Opus 5.5 and GPT-6 Astra. That is the company's account.
If you work in a language other than English, treat the multilingual claim as something to test, not something to assume.
What to do today if your business uses AI
- Don't move anything in production yet. This is a preview and the model will change. Use it to test, not to depend on.
- Test it on your tasks, not the announcement's. Pick five real jobs: answering a customer email, summarizing a contract, writing a product page, fixing a draft, pulling data from an invoice image. Run them through the model you use now and through this one. Compare the results in the language your customers read.
- Work out the cost with your own numbers. At the published prices, 1,000 requests of 2,000 input tokens and 500 output tokens cost about $4.81: $2.72 for input and $2.09 for output. That sum is an estimate from this article, and your real volume will differ.
- Check where your data is processed. Mistral says there will be a European deployment it operates end to end, under European law. If you handle customer data, ask your AI provider which region processes it and on what terms, whatever the model.
- If you want to host it yourself, wait. Neither the license nor the hardware needed has been published. Deciding earlier is deciding blind.
- Keep checking what you publish. A new model does not change the rule: figures, dates and names get verified before a text goes on your website.
Related: OpenAI's GPT-6 Model Guide: Which Model to Pick, and Cost
Sources
- Mistral AI, "Introducing Mistral Large 4", October 6, 2026
- Mistral AI, Mistral Large 4 model card in its documentation
Updates: when Mistral releases the weights, the license or the hardware requirements, it will be added here with a link.