Mistral Large 4 'Le Chonk' Lands in Preview as First Western Open-Weight Model at the 1T-Parameter Tier

Mistral AI opened a public preview of Mistral Large 4 today, a 1-trillion-parameter mixture-of-experts flagship with 49 billion active parameters that the company has already nicknamed "Le Chonk." It's the first Western open-weight contender at the 1T tier, and the claim worth parsing is less about raw size than about geography: the open-weight frontier has, for roughly a year, been a Chinese story.

The headline benchmark is a 59.9% score on AutomationBench, which runs 657 real business workflows across Gmail, Google Sheets, Slack, and Salesforce. That lands ML4 above Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro on Mistral's own chart, though well below Gemini 4 Argon's AutomationBench run and in the general neighborhood of where Claude Opus 5.5 landed in September. Mistral also claims a top-five placement on the Artificial Analysis Cyber Index, including an 82% score on the vulnerability-reproduction-and-patch task and 93% on Cybench. The company attributes part of the cybersecurity lead to closed competitors declining dual-use defensive requests outright, which is a framing choice as much as a technical claim.

On coding, the posture is more candid. ML4 scores 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, while VentureBeat notes the live DeepSWE leaderboard currently places GLM-5.3 and Kimi K3 near 69% and the closed frontier around 74%. CNBC notes the model "still lags behind the frontier in areas such as coding." Mistral's own post describes the reinforcement-learning run as "showing no signs of saturation," which is the sort of thing you say when you want the market to price in a v4.1.

The production story is where European AI policy and compute reality meet. VentureBeat and CNBC report ML4 was trained on 4,000 Nvidia Grace Blackwell GPUs over roughly two months, inside Mistral's own datacenters, across more than 160 languages that include every official EU language. Pierre Stock, VP Science at Mistral, told TechCrunch the compute footprint was "two to three times less than our Chinese competitors."

Preview pricing is $1.36 per million input tokens and $4.18 per million output on the Mistral Studio API. ML4 doesn't yet appear in Artificial Analysis' public evaluations, per VentureBeat, so the independent read-out is pending.

The real event is October 27, when weights and architecture detail are scheduled to drop. Until then, every number here's Mistral's own.

Sources

• https://mistral.ai/news/mistral-large-4/
• https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html
• https://techcrunch.com/2026/10/06/mistrals-new-1t-model-aims-to-leapfrog-closed-and-open-rivals/
• 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://qz.com/mistral-large-4-open-weight-ai-model-launch-100626