Mistral AI opened a public preview of Mistral Large 4, nicknamed Le Chonk, on October 6, giving developers, cybersecurity leaders, and government authorities roughly three weeks to poke at a trillion-parameter mixture-of-experts before the weights land on October 27. Sixteen days of runway. That window, not the launch chart, is where the actual story will be written.

The architecture itself is the headline Mistral wants you to repeat: 1 trillion total parameters, 49 billion active per token, trained in about two months on 4,000 Nvidia Grace Blackwell GPUs inside Mistral's own European data centres, with coverage spanning more than 160 languages including every official EU tongue. VP Science Pierre Stock told TechCrunch the run used "two to three times less" compute than Chinese competitors, a claim that'll either look prescient or embarrassing once the final checkpoint is independently measured.

The benchmark deck is where the humility starts to show through. Mistral's own charts put ML4 at 62% on DeepSWE v1.1, 15% on Harvey's Legal Agent Benchmark, and 67% on Finch finance workflows. On the live DeepSWE board, GLM-5.3 and Kimi K3 sit around 69% under their best configurations. DataCamp's broad aggregate index currently ranks ML4 32nd of 44 models. CNBC noted that the new model "still lags behind the frontier in areas such as coding."

Which is why Oct. 27 matters more than Oct. 6.

Two things flip on that date. First, Artificial Analysis and vals.ai can run the shipped checkpoint instead of a preview endpoint, and the gap between vendor-measured and independently-measured scores becomes legible. Second, the licence text goes public. VentureBeat frames the weights as "expected under a custom Mistral license," which in practice means something adjacent to the Mistral Research Licence rather than Apache 2.0. The distinction isn't academic. A research-only licence lets you benchmark; an Apache-style licence lets you ship a product.

For anyone building a customer-facing AI feature on metered inference, the licence is the number that determines unit economics. Per-token API cost scales linearly against customer acquisition. Self-hosted weights don't. The spread between those two curves is the entire business case for an open-weight frontier-adjacent model from outside China, and it's precisely what Mistral has engineered the Oct. 27 release to answer.

Until then, the preview is a sales deck with latency.

Sources

https://techcrunch.com/2026/10/06/mistrals-new-1t-model-aims-to-leapfrog-closed-and-open-rivals/
https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html
https://venturebeat.com/technology/mistral-debuts-large-4-le-chonk-a-1-trillion-parameter-text-output-model-with-high-benchmarks-planned-for-open-weights-release
https://www.datacamp.com/blog/mistral-4-large-le-chonk
https://www.euronews.com/2026/10/06/what-to-know-about-mistrals-ml4-as-it-bets-on-eu-sovereignty-in-the-us-china-open-weight-a