Mistral has opened a hosted preview of Large 4, a trillion-parameter model whose early benchmark results look competitive, particularly in cybersecurity. The weights are not available yet, so developers can test the service now but will have to wait for the promised later release to run the model themselves.
Mistral AI Watch analysis
What happened
Mistral announced Large 4 on 6 October, according to Kingy AI’s launch analysis. The report says the public preview is available through Mistral Studio, while the company plans to release model weights later in October. That makes this a hosted preview, not an immediately downloadable model.
The report describes a Mixture-of-Experts model with 1.05 trillion parameters in total. It cites Mistral documentation listing 52 billion active parameters, while earlier coverage and Artificial Analysis list 49 billion. Those figures differ, and the report does not resolve the discrepancy.
Artificial Analysis gives the preview an Intelligence Index score of 38 and a Cyber Index score of 50; the report also cites an 82% result on CyberGym-E2E-AA. These are results on named evaluations, not proof that Large 4 will outperform every rival or suit every security team. The report lists standard API prices of $1.36 per million input tokens and $4.18 per million output tokens, with a 50% launch discount for two weeks.
Why it matters
The preview gives developers a chance to test a new European model against real workloads before its weights arrive. That distinction matters: hosted access can support evaluation, but it does not offer the same deployment control as running weights yourself. The eventual licence and operating costs will help determine whether that option is practical.
The early cybersecurity scores make the model worth evaluating, not a ready-made security analyst. Teams will want to check whether it finds genuine issues, explains its evidence and proposes fixes that hold up under review. A benchmark can get a model onto the shortlist; it cannot do the team’s testing for it.
Our read
Large 4 looks like a serious preview with a clear invitation to test, not a verdict delivered by a leaderboard. If you are considering it, try representative tasks on the hosted version and keep results tied to the version you tested. The weights, licence and performance in your own environment are still part of the story.
What to watch
- Whether Mistral releases the weights later in October, and under what licence.
- How the preview performs on developers’ own coding, visual and cybersecurity tasks.
- Whether the differing active-parameter figures and the listed context limits are clarified.
Discussion spark: When evaluating a new model, should strong benchmark results be enough to justify a trial, or should teams wait until they can test the downloadable weights and their own workloads?
Sources and evidence
- Mistral Large 4 Arrives: Europe’s AI Challenger Brings “Le Chonk” to the Fight – Kingy AI (7 October 2026, 23:59 UTC)
not affiliated with or endorsed by Mistral AI