Mistral has released the weights for Mistral Large 4, its trillion-parameter AI model nicknamed Le Chonk, allowing organizations to download and run the system on their own infrastructure. The October 29 release follows a limited preview launched on October 6 and targets software engineering and defensive cybersecurity work.
The French company is making the weights available through Hugging Face, with model access also offered through its API. The release gives businesses and public-sector institutions an alternative to relying entirely on externally hosted AI services, particularly for work involving sensitive source code, security investigations and internal IT systems.
Mistral says additional training since the preview has improved the model’s ability to perform agentic coding tasks, in which an AI system works through a software problem using tools rather than simply generating a standalone code suggestion. Its reported accuracy on Terminal-Bench 4.0 rose from 28% in the preview to 41% for the public release.
The company attributes those gains to three more weeks of reinforcement learning. That training approach uses feedback on attempted tasks to strengthen behaviors that produce successful results. Large 4 also offers low, medium and high reasoning modes, giving users options for how much reasoning effort the model applies to a task.
Mistral positions the updated model as competitive with GLM-5.3, a leading Chinese open-weight rival. It also reports that Large 4 uses 47% fewer tokens per task on average in the comparison. Token consumption matters for inference efficiency, although fewer tokens alone do not establish lower deployment costs: hardware requirements, processing speed and infrastructure utilization also affect the expense of running a model.
Published accounts differ on Large 4’s exact standing against GLM-5.3 on coding benchmarks. The reported Terminal-Bench improvement offers a clearer measure of progress between Mistral’s own preview and public release than a claim to a definitive overall lead. Benchmark results also describe performance under particular test conditions, rather than guaranteeing equivalent results across production software projects.
Cybersecurity is another central focus. Mistral describes Large 4 as a system for detecting, analyzing and responding to threats across software and IT infrastructure. During the preview period, selected partners tested a version with expanded cybersecurity capabilities.
Making model weights available changes who controls deployment. Organizations can retain a local copy rather than depending exclusively on continued access to a provider’s hosted service. For security teams, an internal deployment can also keep the material being analyzed within their own environment. Those benefits still require organizations to supply appropriate computing resources and manage the system’s security themselves.
Open weights are not, by themselves, a complete account of a model’s openness: licensing terms, training-code availability and access to training data are separate considerations. For Large 4’s intended users, the immediate change is practical access to the trained model itself, alongside the option to use Mistral’s API.
Sources: Dealroom, ZDNET, TechCrunch, Fortune