Mistral AI has released Codestral AI for programmers

We present Codestral, our first model for code handling. Codestral is a generative AI model with open-access weights specifically designed for code generation tasks. It helps developers write and interact with code through a unified API for instructions and completions. With its deep understanding of code and English language, this model can be used to create advanced AI applications for software developers.

The 80+ Programming Languages Codestral is trained on a diverse dataset of over 80 programming languages, including the most popular ones such as Python, Java, C, C++, JavaScript, and Bash. It also does well with more specific languages such as Swift and Fortran. This wide range of languages ensures that Codestral can help developers in a variety of environments and projects.

Codestral saves developers time and effort: it can complete code functions, write tests, and complete any partial code using a gap-filling mechanism. Interaction with Codestral will help to improve your programming skills and reduce the risk of bugs and errors.

Setting a new standard for code generation performance Performance. As a 22 billion parameter model, Codestral sets a new performance and speed standard for code generation compared to previous models used for coding.

Get started with Codestral

Download and test Codestral. Codestral is a model with 22 billion parameters and has an open Mistral AI Non-Production License, allowing it to be used for research and testing purposes. Codestral is available for download on HuggingFace.

Use Codestral through its dedicated endpoint. A new endpoint is available with the release: codestral.mistral.ai. This endpoint is preferred by users who use Instruct or Fill-In-the-Middle routes in their IDEs. The API key for this endpoint is managed at a personal level and is not bound by the usual organization restrictions. We offer free use of this endpoint during an 8-week beta period, with access granted on a waitlist to ensure high quality of service. This endpoint is preferred for developers creating plugins for IDEs or applications where clients will use their own API keys.

Create with Codestral on La Plateforme. Codestral is also available on a regular API endpoint: api.mistral.ai, where requests are paid for by the number of tokens. This endpoint and integrations are better suited for research, batch queries, or third-party application development that provide results directly to users without the need for their own API keys.

You can create an account on La Plateforme and start building your applications with Codestral by following this guide. Like all our other models, Codestral is available in our self-deployment offer starting today: contact the sales team.

Communicate with Codestral via le Chat. We provide an instructional version of Codestral that is available today through Le Chat, our free communication interface. Developers can interact with Codestral in a natural and intuitive way using the model's features. We see Codestral as a new step towards enabling everyone to generate and understand code.

Use Codestral in your favorite development and build environments. We worked with community partners to provide access to Codestral in popular tools to improve developer productivity and build AI applications.

Application frameworks. Codestral is integrated into LlamaIndex and LangChain starting today, allowing users to easily build agent-based applications with Codestral.

Integration with VSCode/JetBrains. Continue.dev and Tabnine help developers use Codestral in VSCode and JetBrains environments, now they can generate and interact with code using Codestral.

Here's how the Continue.dev plugin for VSCode can be used to generate code, interactively communicate and edit code in real-time with Codestral, and here's how users can use the Tabnine plugin for VSCode to communicate with Codestral.

For details on how the various integrations with Codestral work, please see our documentation for setup instructions and examples.

Feedback from the developer community

"An autofill model with this combination of speed and quality hasn't existed before, and this will be a game changer for developers everywhere."

- Nate Sesti, CTO and co-founder of Continue.dev

"We are excited about the opportunities Mistral brings and are pleased with the strong focus on coding and development assistance, an area that is key to JetBrains."

- Vladislav Tankov, Head of JetBrains AI department

"We used Codestral to test on our Kotlin-HumanEval benchmark suite and were impressed with the results. For example, in the case of pass rate for T=0.2, Codestral achieved a score of 73.75, outperforming GPT-4-Turbo with a score of 72.05 and GPT-3.5-Turbo with a score of 54.66."

- Mikhail Evtikhiev, researcher at JetBrains

"As a researcher at the company that created the first GenAI developer tool, I have enjoyed integrating the new Mistal code model into our chat product. I am thoroughly impressed with its performance. Despite its relatively compact size, it produces results on par with the much larger models we offer to customers. We tested several key features including code generation, test generation, documentation, customization processes and more. In every case, the model exceeded our expectations. The speed and accuracy of the model will significantly impact the efficiency of our product compared to the previous Mistral model, allowing us to provide fast and accurate assistance to our users. This model stands out as a powerful tool among the models we support, and I highly recommend it to others looking for high quality performance."

- Meital Zilberstein, Head of R&D at Tabnine

"Cody speeds up the internal software development cycle, and developers use features like autocomplete to alleviate some of the day-to-day work involved in writing code. Our internal evaluations show that Mistral's new Codestral significantly reduces the latency of Cody autofill while maintaining the quality of the code offered. This makes it a great choice for autocomplete applications where milliseconds of latency have real value to developers."

- Quinn Slack, CEO and co-founder, Sourcegraph

"I was incredibly impressed with Mistral's new Codestral model for AI-assisted code generation. In my tests, it consistently produced very accurate and functional code, even for complex tasks. For example, when I asked her to complete a non-trivial function to create the new LlamaIndex search engine, she generated code that worked flawlessly despite being based on legacy code."

- Jerry Liu, CEO and co-founder of LlamaIndex

"Code generation is one of the most popular LLM applications, so we are very excited to release Codestral. Our initial tests have shown that it is a great option for code generation workflows due to its speed, favorable context window, and tool support in the instructive version. We tested with LangGraph for self-correcting code generation using Codestral's instructional tools for output, and it worked very well right out of the box (see our video for details)."

- Harrison Chase, CEO and co-founder of LangChain