A new Mistral model can read images, answer questions and work on code, but it cannot yet be downloaded. Mistral opened the public preview of Large 4 on October 6, 2026. The French company promises its weights, the numerical values needed to run the model, by the end of October.
That timing matters for organizations that want to operate AI on their own computers. A hosted preview lets developers test an online service. It does not give them the files needed to inspect or deploy the model themselves.
Large 4 uses a mixture of experts, an architecture that activates only part of the model for each piece of text. Mistral's documentation lists 1.05 trillion parameters in total and 52 billion active parameters. Parameters are the learned numerical values that shape a model's answers.
The company describes Large 4 as multimodal, meaning it works with both text and images. Its documentation also lists tool calling, which lets the model request actions from other software, and structured outputs for applications that need predictable data formats.
Mistral says it trained the model on its own infrastructure in Europe. Its performance comparisons come from the company and should be read as developer-reported results, rather than proof that it leads every competing model.
The preview currently has promotional API prices of $0.68 per million input tokens and $2.09 per million output tokens. Tokens are the small units into which the service splits text. The listed regular rates are twice those amounts.
Testing the preview is already possible. Self-hosting remains a promised release.
