Mistral released the weights for Mistral Large 4 (ML4), its frontier open model, on Hugging Face on Thursday. The company internally refers to the model as Le Chonk and describes it as "the best open agentic coding model in the world."
Three weeks after its initial preview launch, ML4 has undergone substantial refinement through reinforcement learning and testing with customers and partners, according to Mistral. The model outperforms GLM-5.3 on cybersecurity benchmarks and achieves a narrow lead on coding evaluations, the company claims, while operating at a lower cost.
Pierre Stock, Mistral's VP of Science, characterized ML4 as merely "a stepping stone," stating: "We want to be in a position in which we dictate the frontier."
Open weights as essential infrastructure
Stock emphasized Mistral's commitment to open-weight models, describing them as critical for ensuring business continuity, transparency, and auditability. He expressed concern about concentrating AI power in the hands of a few actors: "I'm not comfortable living in a world in which open weights [don't] exist."
Stock referenced recent incidents where OpenAI restricted Cursor's access to its models and the U.S. government ordered Anthropic to suspend Claude Fable 5 and Mythos 5 for foreign nationals. These examples underscore why open weights matter to Mistral's strategy.
Drawing a parallel to essential utilities, Stock explained: "Our stance is that AI is going to be like electricity. You don't want to live in a world in which your neighbor or an oligopoly can cut off electricity from one day to another."
While Anthropic CEO Dario Amodei has argued that open weights alone cannot address AI power concentration due to concentrated access to computing infrastructure, Stock countered that open-weight models can run on trusted infrastructure: "nobody can switch it off because it can run either on your infrastructure or on an infrastructure that you can trust."
Coding and cybersecurity as primary focus areas
Mistral positioned ML4 as "the best open model built in the West," backing this claim with performance metrics from Terminal-Bench 4.0, a component of Artificial Analysis's Coding Agent Index. On this benchmark, Le Chonk scored 41%, compared to Muse Spark Open's 6%, Kimi K3's 21%, and GLM-5.3's 40%.
Beyond raw performance, Stock highlighted efficiency gains: "What the graph doesn't tell you is that it's more efficient than the competitors. Le Chonk uses 47% fewer tokens per task on average than GLM 5.3. It means that you can get things done way faster and at a lower cost."
Stock identified coding and cybersecurity as the two areas where ML4 delivers the most customer value. Mistral invested specifically in strengthening the model's defensive cybersecurity capabilities.
On the CyberGym E2E benchmark from the Artificial Analysis Cyber Index, which evaluates defensive cybersecurity capabilities, ML4 scored 79% compared to GLM-5.3's 29%. The model also exceeded DeepSeek v4.1 Flash 1017's 76% and matched MiMo-V2.6-Pro's 79%.
Stock described how some Mistral customers are already deploying the model for automated penetration testing, where specialized agents analyze application logs and identify potential security vulnerabilities.
Rapid gains with constrained resources
Mistral trained ML4 over approximately two months using roughly 3,800 Nvidia Grace Blackwell GPUs in its European data centers before the October preview launch. In the three weeks following that release, Stock attributed "significant improvement" to reinforcement learning efforts, with the Terminal-Bench 4.0 score rising from 28% to 41%.
Stock claimed that Mistral completed this reinforcement learning phase using only 10 megawatts of power, "a fraction of what all the competitors have."
When asked how Mistral can assert cost efficiency with certainty, Stock replied: "I have no information about someone being able to do it with this amount of resources."
Stock suggested that Mistral's resource efficiency could translate into further advances with additional capital: "If you give us 10 times more capital, then we'll factor in these efficiency gains."
Roadmap and broader ambitions
Throughout the briefing, Stock consistently referred to ML4 as "a stepping stone," signaling Mistral's plans for future development.
Mistral confirmed that smaller versions of Le Chonk will arrive in the coming months. Stock stated: "There's going to be a family of models, and one of them is going to be on a laptop."
The company identified potential applications for ML4 in modernizing legacy codebases, finance, manufacturing, industrial engineering, and geospatial work. Mistral has recently made notable progress in COBOL migration, an area where OpenAI has also demonstrated capability.
Looking forward, Mistral articulated a broader strategic vision: "We think the only way forward [is] to own the entire stack."