Mistral has published the weights of Mistral Large 4, the roughly trillion-parameter model it previewed in early October. The version now on Hugging Face is noticeably stronger: three more weeks of reinforcement learning, guided by customer feedback, have put it at the top of open models for agentic coding and cyber defence, according to the Paris company.
Nicknamed "Le Chonk", the model reads text and images and produces text and code, activating about 52 billion of its parameters for each task. Alongside the full model, Mistral is releasing a quantised version, a compressed one that needs less memory, which reportedly runs on four Nvidia B200 or B300 GPUs.
Image: Mistral AI — https://mistral.ai
The clearest gain is in coding. On Terminal-Bench 4.0, which tests agentic coding in a terminal, ML4 rose from 28% at preview to 41%, just ahead of GLM-5.3 at 40%. According to Artificial Analysis, on its Coding Agent Index the model rivals GLM-5.3 while using 47% fewer tokens per task on average, a saving that adds up when agents run for hours.
In cyber defence, ML4 scored 79% on CyberGym-E2E, a test of whether a model can find, reproduce and patch memory-safety bugs in open-source C and C++ projects. That matches MiMo-V2.6-Pro and beats DeepSeek V4.1 Flash, at 76%. Some customers already run automated penetration tests with specialised agents that probe their own applications for weaknesses.
Early testers, whom Mistral describes only as governments and enterprises, wanted more control over how long the model reasons. The company added low and medium reasoning modes next to high, and made high 50% more efficient.
Pierre Stock, Mistral's VP of Science, treats the open release as a question of continuity. "AI is going to be like electricity", he said, and no company should depend on someone who can cut its access "from one day to another". Once weights circulate online, he argues, access cannot be withdrawn, though rivals dispute the safety of open models. He applies the same logic to the Chinese labs that dominate open models: nothing guarantees they will keep publishing weights, and their models offer less transparency and could become attack vectors.
Mistral says it built ML4 with a fraction of the resources of closed labs and Chinese rivals, a claim hard to test given how little those labs disclose about training costs. The model was trained on 3,800 Nvidia Grace Blackwell GPUs in Europe, while reinforcement learning ran on the company's own compute in France, drawing about 10 megawatts. "Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion", said CEO Arthur Mensch.
Smaller, distilled versions will follow in the coming months, one of them running on a laptop, but Mistral has not committed to dates. API pricing for ML4 has not been published; Stock argues that per-token rates mislead when a model needs far fewer tokens per job.
ML4 is available on Hugging Face, through Mistral's API and to some Vibe users. The harder test comes now: Stock wants Mistral to "create the cadence of the frontier", not reach it just once.
No comments yet. Be the first!