French AI company Mistral released the full weights for its Mistral Large 4 (ML4) model on Thursday, calling it "the best open agentic coding model in the world" and declaring it superior to competitors on both cybersecurity and coding benchmarks. Available in preview just three weeks earlier, the model—nicknamed Le Chonk by Mistral—has advanced substantially since then through reinforcement learning and customer testing, the company says. Mistral VP of Science Pierre Stock described the release as merely "a stepping stone," signaling the French startup's ambition to lead the AI frontier.
On Terminal-Bench 4.0, part of Artificial Analysis's Coding Agent Index, Le Chonk achieved a 41% score, narrowly surpassing GLM 5.3's 40% and far outpacing Muse Spark Open's 6% and Kimi K3's 21%, according to Mistral's figures. The company says the model uses 47% fewer tokens per task on average than GLM 5.3, making it faster and cheaper to run. In CyberGym E2E from the Artificial Analysis Cyber Index, which evaluates defensive cybersecurity abilities, ML4 scored 79%—matching MiMo-V2.6-Pro and exceeding DeepSeek v4.1 Flash 1017's 76% and GLM-5.3's 29%. Mistral's own Terminal-Bench 4.0 score climbed from 28% at preview to 41% in three weeks.
Stock told reporters in an embargoed briefing that Mistral conducted its reinforcement learning on just 10 megawatts of power, which he characterized as "a fraction of what all the competitors have." The initial training ran for roughly two months on approximately 3,800 Nvidia Grace Blackwell GPUs housed in European data centers. According to Stock, Mistral fully backs "the power of open weight," calling it vital for business continuity, transparency, and auditability. "I'm not comfortable living in a world in which open weights [don't] exist," he said, adding that Mistral's philosophy holds AI should function like electricity—something no neighbor or oligopoly can shut off overnight.
Stock explained that Mistral targeted coding and cybersecurity as the two areas with the greatest potential to transform customer businesses through AI. Some clients are already deploying the model for automated penetration testing, where specialized agents examine application logs and probe for vulnerabilities. The company says ML4 can help software engineers modernize legacy codebases and has applications in finance, manufacturing, industrial engineering, and geospatial work. Stock confirmed smaller versions of Le Chonk will ship in coming months, including one designed to run on a laptop. When asked how Mistral knows it's more cost-efficient than competitors, Stock responded, "I have no information about someone being able to do it with this amount of resources," and added that if the company received ten times more capital, it would multiply those efficiency gains. The French startup frames its resource efficiency as proof that concentration of AI power isn't inevitable—even as critics like Anthropic CEO Dario Amodei argue that computing infrastructure remains the real bottleneck, not model weights alone. For companies weighing build-versus-buy decisions on AI tooling, the tension between proprietary scale and open adaptability will sharpen as both camps race to prove their approach can deliver enterprise-grade reliability at speed.