Depend onMicrosoftSupported by a $6 billion valuationFranceAIStartupsMistralRecently released its first generative artificial intelligence model for coding——Codestral.
The model is designed to help developers write and interact with code, supporting more than 80 programming languages including Python, Java, C++ and JavaScript. Mistral said in its blog post that Codestral can complete coding functions, write tests, fill in parts of the code, and answer questions about the code base in English.
Although Mistral describes Codestral as “open,” its license prohibits the use of the model and its output for any commercial activity, with the exception of “development.” The license specifically prohibits “any internal use by employees in the course of company business activities.” This is likely due to the fact that some of Codestral’s training data contains copyrighted content.
It is reported that the Codestral model has 22 billion parameters and requires a powerful PC to run. The model beats competitors in some benchmarks, but the reliability of these benchmarks is still in doubt. Therefore, the actual value and practicality of Codestral remains to be seen.
According to a Stack Overflow survey in June 2023, 44% of developers said they use AI tools in their development process, and 26% of developers plan to use them soon. However, these tools are not without flaws. GitClear analyzed more than 150 million lines of code submitted to project repositories in the past few years and found that generative AI development tools led to more wrong code being pushed into the code base.
Additionally, a Purdue University study found that OpenAI’s ChatGPT got more than half of its responses to programming questions wrong.
This morning, Mistral launched a hosted version of Codestral on its Le Chat conversational AI platform and its paid API, with plans to build it into application frameworks and development environments such as LlamaIndex, LangChain, Continue.dev, and Tabnine.