Mistral AI releases Mistral Large and launches Le Chat

Mistral AI released Mistral Large, a closed-weight model it described as its most capable, alongside Le Chat, a ChatGPT-style assistant built on top of it. The company said Mistral Large trailed only GPT-4 among models generally available through an API, citing results on reasoning, multilingual, coding and maths benchmarks where it outperformed Claude 2, Gemini Pro 1.0 and Meta’s LLaMA 2 70B. The model handled English, French, Spanish, German and Italian natively, took a 32K-token context window, and supported function calling and a JSON output mode.

The releases were notable chiefly for what they were not: Mistral had built its reputation on openly licensed models such as Mixtral, and Mistral Large broke with that pattern by shipping closed, available only through Mistral’s own API and, from the same day, through Microsoft Azure. Mistral said the model remained available for self-deployment for customers with strict data-control requirements, but its flagship product joined the closed-weight field it had previously positioned itself against.

Microsoft announced a partnership alongside the launch, investing $16 million in Mistral and adding its models — including Mistral Large — to Azure AI Studio and Azure Machine Learning. For Microsoft, already OpenAI’s largest backer, the deal gave it a second frontier-model supplier without building one in-house; for Mistral, not yet a year old, it brought distribution and cloud credibility it could not otherwise match. Mistral Small, a smaller model tuned for latency-sensitive workloads, launched the same day.

The episode became an early data point in the argument over what “open” meant for a company that had used it as a selling point, a debate that continued as European rivals weighed commercial pressure against the open-release strategy that had built their audience.