Codestral

What is Codestral?

Codestral is Mistral AI's first major open-weight language model (LLM) specifically designed for code generation tasks. Released in mid-2024, this 22-billion-parameter (22B) model was exhaustively trained on a dataset containing more than 80 programming languages. It is designed to excel in Fill-in-the-Middle (FIM) code completion, writing software tests, and following natural language programming prompts with high performance and low latency.

How it works

Codestral works by receiving unfinished code sequences or instructions described by the user. Thanks to its 22B parameter architecture, it can fill in missing code snippets while maintaining structural coherence. The model supports a wide context window of 32k tokens, meaning it can read and understand long files and entire code dependencies at once before suggesting refactoring solutions or generating new programming functions.

Key features

• 32k Token Context Window: Ability to analyze repositories and multiple files at once to avoid out-of-scope code suggestions.
• 80+ Languages Support: High proficiency in common languages (such as Python, JavaScript, TypeScript, C++, Java, and Go) as well as more specific languages (such as Fortran, Swift, and Cobol).
• Fill-in-the-Middle (FIM): Excellent performance in autocomplete tasks in the middle of a file, rather than just suggesting lines at the end.
• Lightweight Execution: Optimized model to provide fast responses via API or to be run locally by developers with compatible commercial GPUs.

Available integrations

As a foundation model, Codestral can be coupled with various development tools via its APIs:

• Popular extensions such as Continue.dev, Llama.coder, and Tabby.
• Code editors compatible with OpenAI/Mistral APIs (such as VS Code and Cursor).
• Local enterprise environments via Ollama.

Who it is for

• Software Developers looking for a very fast code copilot for autocomplete and test generation.
• Technology Companies wanting to host their own programming AI models on private servers for compliance reasons.
• Development Tool Creators interested in building AI assistants based on a reliable programming model.

Pricing

Pros and cons

Pros:

• Excellent API cost-benefit ratio compared to giant language models like GPT-4.
• Remarkable proficiency in Fill-in-the-Middle (FIM) autocomplete.
• Generous 32k context window that accommodates large chunks of code.

Cons:

• The model usage license restricts direct commercial use without prior subscription to Mistral's commercial APIs.
• Requires advanced enterprise-grade graphics cards (GPUs) to run locally with professional performance.

Alternatives to Codestral

Other open and closed-source AI models focused on programming include: DeepSeek Coder V2, CodeLlama (from Meta), StarCoder 2, and GitHub Copilot (OpenAI Codex).