Mistral Large 4 in Pre-release, Weights Unveiled on October 26

Mistral AI opens access to Mistral Large 4 via API and plans to release its weights on October 26, along with quantized variants and an announced starting price. The model, the first major launch since Large 3, combines instruction, reasoning, and vision, and is positioned as an open weight response for sectors facing strict hosting constraints.

Open Pre-release, Controlled Access, and Weight Publication Announcement

Mistral Large 4 is accessible in public pre-release via an API, under the identifier mistral-large-4. At the same time, a less moderated private pre-release allows cybersecurity partners and public authorities to access the model's weights. The publication of the weights is scheduled for October 26, with quantized versions in FP8 and FP4, deployable on four to eight B200 or B300 GPUs. The displayed price is $1.36 per million tokens for input and $4.18 for output. The documentation currently mentions a discount that halves these prices, with no indication of duration. Mistral Large 4 is also set to become the default model for the Vibe assistant, which currently uses Mistral Medium 3.5.

MoE Architecture, Integrated Vision, and Fine-tuning of Reasoning

As the first major model since Large 3, Mistral Large 4, nicknamed "the Chonk," combines the instruct and reasoning modes, as well as agentic capabilities, into a single version. The reasoning_effort parameter allows for adjustment of the degree of reasoning, similar to Mistral Small 4. This model uses a Mixture-of-Experts architecture totaling 1.05 trillion parameters, of which 49 billion are utilized, and also includes a vision encoder with 1.6 billion parameters for image analysis alongside text. Its context window capacity is one million tokens and it supports over 160 languages, covering all official languages of the European Union. Mistral AI positions itself at the cutting edge among open models for sectors such as cybersecurity, finance, industrial production, and multimodal applications, claiming to outperform closed models for object detection in images (visual grounding).

Open Weight Positioning, Chinese Competition, and Sovereignty

After a period where Mistral AI was perceived as lagging in rankings and more focused on deploying solutions for its clients, the company claims it wants to demonstrate its capabilities with Mistral Large 4, an open weight model. Mistral AI presents this model as the best open weight model designed in Europe or the United States and states that it ranks among the best open models in the world according to aggregated rankings. However, in the realm of open weights, Chinese labs like Kimi, DeepSeek, and GLM dominate, and Mistral AI acknowledges that several of these models remain superior in certain areas. To justify the choice of open weights, Mistral AI highlights the ability for companies to host the model internally, a common requirement in sectors like finance, healthcare, or defense, as well as independent access from a provider, whereas some leading closed models are currently restricted for cybersecurity. The argument of sovereignty is also put forward: Mistral Large 4 was entirely trained in Europe on 4,000 GPUs in the company's data centers. This computing power is described as modest compared to American giants, but Mistral AI plans to increase it through a €3 billion fundraising completed in September.