Mistral released the weights of Mistral Large 4, a 1 trillion-parameter open model aimed at coding and cyber defense

Mistral has released the open weights of Mistral Large 4, the 1 trillion parameter multimodal model it nicknames Le Chonk. The model went into public preview earlier this month. Mistral says three more weeks of reinforcement learning have since improved its coding and defensive cybersecurity results, and that it now ranks as the best open weight model for coding.

The numbers come from Mistral's own announcement. On Cognition's private FrontierCode 2.0 benchmark it reports a lead at 19.7%. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it rose 13 points from the preview to 41%, with 30% fewer output tokens. ZDNET notes that GLM-5.3 still edges it by one point on the overall index, 49 to 48, and cautions that benchmark scores deserve a grain of salt.

Cybersecurity is the second pitch. Mistral reports 51% on the Artificial Analysis Cyber Index and an average refusal rate of 96.6% on malicious cyber requests, which it calls the highest it measures among open models. It also says it saw potential for uplift in real attacks during the preview and is tightening its guardrails. Enterprise design partners such as Orange Cyberdefense get preview access to a specialized variant for penetration testing and red teaming, under vetting and contractual limits.

The model has three reasoning modes: Low, Medium and High, with High as the default. Mistral says it was trained from scratch on about 4,000 Nvidia Grace Blackwell GPUs over two months in its own European data centers. Yahoo Tech, citing the preview launch, reported roughly 1.05 trillion parameters with 52 billion active, a context window of up to 1 million tokens and published API pricing of $1.36 per million input tokens and $4.18 per million output tokens, with preview discounts.

For developers, the draw is control. Open weights mean a team can self host, fine tune on its own security data through Mistral Forge, or run it under European law, and Mistral frames this as sovereign AI. The token efficiency claim matters if you run coding agents at volume, since cost per task often matters more than a one point benchmark gap. Mistral also says it stays ahead of GLM-5.3 and Kimi K3 in blind human preference tests run by Surge, behind only Claude Opus 5.5 among the models evaluated.

Treat the results as vendor claims until independent tests land. A model this size also needs serious hardware to run locally, and Yahoo Tech noted the release is subject to the model's license. If you build coding agents, the practical step is to try the Low and Medium modes on your own repositories and compare cost per merged change against what you use today. Mistral says its training gains show no signs of saturation, so expect more updates.

Building with AI? See this week's hot AI repos and how to use each one.