Mistral, the Paris-based AI company, has published the weights of Mistral Large 4, a model with about one trillion parameters that the company nicknames “Le Chonk.” The release came on October 29, roughly three weeks after Mistral first showed the model as a preview in early October. Readers who followed our earlier coverage of that preview will notice a different story this time, because the version now available for download performs noticeably better than the one shown at launch, according to the company’s results.

For anyone new to the term, a model’s weights are the numbers it learned during training. Publishing them means companies, governments, and researchers can download the model and run it on servers they control, instead of reaching it through someone else’s service. Mistral put the files on Hugging Face, the main public hub for models of this kind.

Three extra weeks of reinforcement learning

The gap between the preview and the final release comes from additional training. According to The Next Web, Mistral spent three more weeks on reinforcement learning, a method where a model improves by being rewarded each time it solves a task well, and feedback from customers guided the process.

The clearest result appears on Terminal-Bench 4.0, a test of how well a model handles coding tasks inside a terminal. Mistral Large 4 climbed from 28% at preview to 41%. In the chart Mistral shared, that places it just ahead of GLM-5.3 at 40%, with DeepSeek V4.1 Flash at 31% and Kimi K3 at 21%. These figures were published by Mistral, so independent testing will show how they hold up.

Fewer tokens for the same kind of work

On Artificial Analysis’ Coding Agent Index, Mistral says Large 4 performs at a level comparable to GLM-5.3 while using 47% fewer tokens per task on average. Tokens are the small pieces of text a model reads and writes, and they set both the cost and the duration of a job. A coding agent that runs for hours can burn through a huge number of them, so a saving of that size adds up quickly. Mistral also claims the model ranks above every other open model on that index, a point TechCrunch reports as the company’s statement.

The cybersecurity results follow a similar pattern. On CyberGym-E2E, a test of whether a model can find, reproduce, and patch real memory-safety bugs in open-source C and C++ projects, Mistral Large 4 scored 79%. That matches MiMo-V2.6-Pro and sits ahead of DeepSeek V4.1 Flash at 76%.

Reasoning modes and security work

Early testers, whom Mistral describes only as governments and enterprises, asked for more control over how long the model spends thinking before it answers. Mistral responded by adding low and medium reasoning modes next to the existing high mode, and it made the high mode 50% more efficient.

Some customers already use the model for automated penetration testing, where specialized agents probe an application for weak points and write up what they find.

Size, hardware and license

Large 4 is a mixture-of-experts model, which means only part of the network switches on for each task. Reports put the active share at roughly 49 to 52 billion parameters, out of about one trillion in total. The model accepts text and images and produces text and code, and Hugging Face’s description lists a context window of one million tokens, enough room for very long documents or large code projects.

Mistral also released a quantized version, a compressed edition that needs less memory, and it runs on four Nvidia B200 or B300 GPUs. That is still data center hardware, not something for a home computer.

The license is the Mistral Ecosystem Open-Weight License, described by Hugging Face as a variant of the MIT license with limits on model-as-a-service offerings. Anyone planning to build a commercial service on top of the model should read those terms carefully.

Mistral’s argument for open weights

Pierre Stock, Mistral’s vice president of science, framed the release around continuity. “AI is going to be like electricity,” he told The Next Web, and he argued that no company should depend on a supplier that can end its access overnight. Once the weights are public, he said, nobody can withdraw them. Speaking to TechCrunch, he added that defenders should receive strong cyber capabilities as early as possible, since malicious actors may use the same technology to attack.

Critics read the same fact differently. Because published weights cannot be recalled, some rivals argue that open models carry more risk, a view The Next Web also reports. The debate has no settled answer, and this release gives both sides a large new example.

Training in Europe on a smaller budget

Large 4 was trained on 3,800 Nvidia Grace Blackwell GPUs in Europe, and the reinforcement learning stage ran on Mistral’s computing resources in France, drawing about 10 megawatts, according to The Next Web. CEO Arthur Mensch said other labs have raised hundreds of billions to develop comparable capabilities, while Mistral has raised roughly $6 billion in three years. The Next Web adds that such comparisons are hard to test, because some labs disclose little about their training costs.

What comes next

Smaller distilled versions will follow in the coming months, including one that runs on a laptop, though Mistral has not committed to dates. A distilled model is a smaller model trained to imitate a larger one, which makes it cheaper to run. API pricing for Large 4 has not been published, and Stock argued that per-token rates can mislead when a model needs far fewer tokens to finish a job. He also said he wants Mistral to keep releasing at the frontier pace rather than reaching it once, so the next releases will show whether that ambition holds.