Mistral AI has unveiled Codestral, its first generative AI model designed for code generation. This open-weight model aims to assist developers by offering extensive support across a wide array of programming languages and enriched performance benchmarks.
Codestral is proficient in over 80 programming languages, ranging from popular ones like Python, Java, and JavaScript to more specialized languages such as Swift and Fortran. This extensive language support ensures that developers working on diverse projects can benefit from Codestral’s capabilities. The model can complete coding functions, write tests, and fill in partial code, thereby reducing the risk of errors and enhancing productivity.
The model sets a new standard in code generation performance. At 22 billion parameters, Codestral offers unparalleled performance and low latency, significantly outperforming previous models. Tests show it excels in several benchmarks, including HumanEval for Python, CruxEval for Python output prediction, and Spider for SQL performance. Its fill-in-the-middle capabilities were notably efficient, even compared to larger models like DeepSeek Coder 33B.
To get started, developers can download and test Codestral under the Mistral AI Non-Production License. The model is available on HuggingFace for research and testing purposes. For commercial use, licenses are available upon request. Codestral also offers dedicated endpoints for different purposes. The new endpoint, codestral.mistral.ai, is ideal for developers working within their IDEs and is free during an eight-week beta period. For other applications, Codestral is accessible via the api.mistral.ai endpoint, where queries are billed per token.
Developers can also interact with Codestral through Le Chat, Mistral AI’s conversational interface. This feature aims to make it easy for developers to leverage the model’s capabilities intuitively. Codestral integrates seamlessly with popular developer tools and environments, such as VSCode and JetBrains, thanks to partnerships with Continue.dev and Tabnine.
The developer community has already expressed excitement about Codestral’s potential. Nate Sesti of Continue.dev highlighted its unprecedented combination of speed and quality. Vladislav Tankov of JetBrains praised Mistral’s focus on development assistance, while Mikhail Evtikhiev noted its superior performance on the Kotlin-HumanEval benchmark. Tabnine’s Meital Zilberstein lauded its speed and accuracy, which significantly enhance their product’s efficiency. Quinn Slack from Sourcegraph emphasized the reduction in latency for autocomplete tasks, and Jerry Liu from LlamaIndex commended its functional code generation for complex tasks. Harrison Chase from LangChain found it highly suitable for self-corrective code generation workflows.
With Codestral, Mistral AI aims to empower developers and democratize coding. As it gains traction, the model is poised to become an invaluable tool for software development, offering reliable and advanced AI-powered code generation.
You can read the original article here: https://mistral.ai/news/codestral/
