Mistral AI released a new AI model for code generation

French AI startup Mistral has launched its first generative AI model for code generation. The new product is called Codestral.

Like other analogues, the tool is designed to help developers write and edit code when creating software. As follows from the announcement on the startup’s blog, the model was trained to work with more than 80 programming languages, including Python, Java, C++ and JavaScript. In particular, Codestral can write code and tests, fill in gaps in the code, and answer questions about the codebase in English.

Mistral classifies the model as open-source, but the startup’s license prohibits the use of Codestral and products created with its help for commercial purposes, including for performing work tasks within any company. According to unconfirmed information from the TechCrunch portal, this may be due to the fact that the startup partially trained the model using copyrighted context. The tool can be used privately for research purposes or code testing.

The model is quite “heavy” and has 22 billion parameters, which is why its deployment is currently only possible on computers with high computing power. At the same time, such a number of parameters allows it to cope with tasks more efficiently and surpass competitors in certain criteria.

As TechCrunch writes, the model may be impractical for most developers, since it requires the use of powerful hardware, but it may serve as a new basis for discussion about the feasibility of using AI code generation models as a programming aid. In particular, last year’s survey on Stack Overflow revealed that about 45% of developers already use such auxiliary tools, and another 25% plan to use them in the future. At the same time, the quality of work of such models cannot yet be called ideal – they make many mistakes (sometimes serious) and do not always give the correct answers.

The startup published the Codestral model for free on Hugging Face and on its AI platform Le Chat, and also provided a paid API. The company reported that work is currently underway to integrate the model into the frameworks of some applications and development platforms.