Mistral has launched Codestral, a new open-weight AI model designed to help developers write and interact with code. Codestral is a 22B parameter model with a 32K context window. It was trained on over 80 programming languages, including Swift, Bash, Python, Java, C++, and JavaScript.
The model can be downloaded on HuggingFace and is available through Mistral's API platform. Developers can also use Codestral for free on Mistral's conversational AI platform, Le Chat.
Codestral sets a new standard for code generation performance, particularly in the context of latency. When compared to existing code-specific models, Codestral delivers impressive results in Python and SQL benchmarks.
The model is impressively fast—whether using the API or running completely offline. See this side-by-side by Nick Dobos of Codestral with Mistral 7b.
Codestral @MistralAILabs first impression:
— Nick Dobos (@NickADobos) May 29, 2024
1. 80 languages is crazy. Finally someone included Swift. Which a lot of OS models skip
2. Really fucking fast. wtf.
It’s a 22b model and it’s significantly faster than mistral 7b. Are they using groq to serve it?? Comparison: pic.twitter.com/dAFaporiRd
Mistral has worked with community partners to integrate Codestral into popular developer tools and frameworks, including LlamaIndex, LangChain, Continue.dev, and Tabnine. This allows developers to use Codestral within familiar environments like VSCode and JetBrains. Early feedback from developers highlights Codestral's speed, accuracy, and potential to revolutionize the coding process.
It's important to note that Codestral is licensed under their new Mistral AI Non-Production License, which restricts its use to research and testing purposes only. This means that users cannot utilize the model or its outputs for commercial activities, including internal usage by employees within a company's business context.