Mistral unveils Large 4 “Le Chonk”, its multimodal model with a trillion parameters

On October 6, 2026, Mistral AI announced in an online broadcast the availability of a public preview of its new model, Large 4, internally known as Le Chonk. The information comes from a company press release and was picked up the same day by VentureBeat and Reuters. According to the published data, the model contains a total of one trillion parameters and maintains 49 billion active weights during inference, placing it among the largest AI systems currently available on the European market. The announcement highlights that this scale puts the model in the same class as the biggest offerings from U.S. and Chinese giants, but with a distinctly European focus on data sovereignty.

Technical details of the model

Le Chonk was trained from scratch for roughly two months, using 4 000 Nvidia Grace Blackwell GPUs distributed across Mistral’s own data centers in Europe. The architecture is based on a Mixture‑of‑Experts (MoE) approach, which allows only a fraction of the parameters—49 billion—to be activated at each inference step. This selective activation reduces computational and energy requirements compared with a dense model of comparable size. Mistral describes the infrastructure as “sovereign,” designed to operate under zero‑data‑retention policies, making it attractive to customers subject to strict privacy and data‑protection regulations.

Linguistic coverage and multimodal capabilities

The model has been trained to understand and generate text in more than 160 languages, including the 24 official languages of the European Union. In addition, it incorporates visual capabilities that enable it to process images and combine visual information with text, opening possibilities for tasks such as object recognition, analysis of scanned documents, and generation of visual descriptions. Mistral emphasizes that the combination of a massive parameter count and the MoE architecture provides superior efficiency on tasks that require multimodal reasoning, i.e., the ability to reason simultaneously over textual and visual data.

Launch plan and availability

Mistral announced that the model’s API will be available to developers starting October 27, 2026, after a three‑week testing phase that includes cybersecurity experts, regulatory authorities, and a select community of developers. The company also pledged to release the model weights on the same date under its own license, with the goal of fostering open research and adoption in sovereign environments. This dual release strategy—both an API and downloadable weights—aims to encourage broad uptake while respecting data‑sovereignty requirements.

• Programming
• Cybersecurity
• Finance
• Manufacturing
• Visual understanding

License and business model

Le Chonk’s weights will be distributed under a proprietary Mistral license, the specifics of which have not been publicly detailed. The company has not announced a price for API access, stating that commercial information will be disclosed in a later phase. Mistral stresses that its model is intended to be deployed on sovereign infrastructures, offering zero‑data‑retention options and compliance with local regulations, distinguishing it from offerings that rely on international public clouds.

Performance comparisons and limitations

In the announcement, Mistral cited benchmark results that allegedly outperform some Chinese models in specific areas, as reported by Reuters. However, the company also clarified that those comparisons come from its own internal testing and that there are no uniform third‑party data sets that would allow independent validation of the model’s superiority. Moreover, the MoE architecture means performance can vary depending on workload and the availability of activated experts, representing a limitation to consider for critical applications where consistent performance is essential.

With the release of Le Chonk, Mistral reinforces its position as one of the few European players capable of offering a multimodal model at a scale comparable to U.S. and Chinese giants, while adopting a focus on data sovereignty and deployment flexibility. If the developer community quickly adopts the API and open weights, the model could accelerate the creation of local applications in sectors such as programming, cybersecurity, and manufacturing, reducing reliance on external services and boosting innovation within stricter regulatory frameworks.