Mistral Large 4 opens for testing before its weights are released
Mistral has opened a public API preview of Large 4, a multimodal model also called Le Chonk. Downloadable weights are planned for the end of October, after testing with cybersecurity specialists, selected partners and public authorities. The staged release gives developers hosted access now while the French lab continues safety work and training before enabling deployment on other infrastructure.
Artificial Intelligence··Night
Large 4 starts with hosted preview access
Mistral, the French AI laboratory, opened a public preview of Mistral Large 4 on October 6. Also called Le Chonk, it is a multimodal model, meaning it works with visual information as well as text. Developers can try it through the Mistral Studio application programming interface, the connection used by software to send requests to a model. Downloadable model weights, the numerical settings learned during training, are planned for the end of October. Current access runs through the company’s hosted service with safety guardrails.[1], [2]
Mistral describes Large 4 as having 1 trillion total parameters and 49 billion active parameters. These counts concern the model’s learned numerical structure. The announcement says a substantial part of its training data was multilingual, covering more than 160 languages including all official European Union languages. The company is continuing reinforcement learning during the preview period, so the release begins while further training work is still underway.[1]
Safety testing precedes downloadable weights
Before publishing the weights, Mistral says it is testing the model in real settings with cybersecurity specialists, vetted partners and public authorities. Those participants are to use the same model with reduced moderation and expanded cyber capabilities. Pierre Stock, Mistral’s vice president of science, describes the aim as supporting defensive use while limiting malicious attacks. The staged timetable puts hosted access and the more widely deployable weights at different points in the release process.[1], [2]
Enterprise uses shape the model’s next steps
The proposed work includes cybersecurity, finance and chip design. Mistral also gives visual examples such as examining engineering drawings, finding evidence in PDF documents and locating objects in large satellite images. Those examples describe intended applications rather than demonstrated improvements for every customer. Stock argues that open weights make a model easier to examine; that assessment is the executive’s position. The company’s core audience includes enterprises and public institutions with security concerns.[1], [2]
Mistral says training took place on its own European computing infrastructure, which also serves the preview. Its plan includes later operation in private clouds or on an organization’s own premises. More architectural detail and information about post-training are scheduled to accompany the weight release. Large 4 is also intended to underpin future specialized Mistral models, extending the laboratory’s own model development alongside the inference services it supplies.[1]
