Mistral Releases Mistral Large 4 Open Weights, Touts Coding and Cyber Gains

Mistral released the weights for Mistral Large 4 on Hugging Face on Thursday, saying the model improved significantly since its preview three weeks ago through reinforcement learning and customer testing. The company claims it is the best open agentic coding model in the world, beating GLM-5.3 on cybersecurity and coding benchmarks at a fraction of the cost.

Open weights and the frontier

In an embargoed press briefing attended by The New Stack, Pierre Stock, Mistral's VP of Science, called Mistral Large 4 "a stepping stone" and said, "We want to be in a position in which we dictate the frontier." Stock said Mistral fully supports "the power of open weight," calling it essential to business continuity, transparency, and auditability. "I'm not comfortable living in a world in which open weights [don't] exist," he said. He compared AI to electricity, saying, "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." Recalling when OpenAI announced plans to cut off Cursor's direct access to its models and when the US government ordered Anthropic to suspend Claude Fable 5 and Mythos 5 for all foreign nationals, Stock argued that open weights mean "nobody can switch it off because it can run either on your infrastructure or on an infrastructure that you can trust." Anthropic CEO Dario Amodei has argued that open weights alone cannot resolve the concentration of AI power because access to computing infrastructure remains concentrated.

Coding and cybersecurity benchmarks

Mistral calls Mistral Large 4 "the best open model built in the West." On Terminal-Bench 4.0, part of Artificial Analysis's Coding Agent Index, the company says Le Chonk, its nickname for the model, ranks as the top open model for agentic coding with 41%, compared with Muse Spark Open's 6%, Kimi K3's 21%, and GLM 5.3's 40%. Stock said Le Chonk uses 47% fewer tokens per task on average than GLM 5.3. Mistral also said coding and cyber are two main areas that can help customers transform their business with AI. In CyberGym E2E from the Artificial Analysis Cyber Index, which tests defensive cybersecurity capabilities, Mistral Large 4 scored 79%, compared with GLM-5.3's 29%, DeepSeek v4.1 Flash 1017's 76%, and a tie with MiMo-V2.6-Pro's 79%. The chart also showed Z.ai's smaller GLM-5.3 Flash at 74%. Mistral said some customers use the model to support automated pen testing, where specialized agents analyze an application's logs and test for potential entry points.

Efficiency and next steps

Mistral launched Mistral Large 4 in public preview at the start of October after about two months of training on about 3,800 Nvidia Grace Blackwell GPUs in its European data centers. In the last three weeks, Stock said the model underwent significant improvement from reinforcement learning, and Mistral's Terminal-Bench 4.0 score rose from 28% at preview to 41%. He said the French company ran that reinforcement learning on only 10 megawatts of power, "a fraction of what all the competitors have," and said, "I have no information about someone being able to do it with this amount of resources." Stock also said Mistral could advance further with the right investments: "If you give us 10 times more capital, then we'll factor in these efficiency gains." Developers can expect smaller versions of Le Chonk in the coming months, with Stock confirming, "There's going to be a family of models, and one of them is going to be on a laptop." Mistral says Mistral Large 4 can help software engineers modernize legacy codebases, as well as in finance, manufacturing, industrial engineering, and geospatial applications. Looking ahead, Stock said, "We think the only way forward [is] to own the entire stack."

Sources: The New Stack · The Decoder