Codestral
By Mistral AI
Codestral is Mistral AI's family of code-specialized large language models, first released in May 2024, trained on over 80 programming languages and optimized for code generation, completion, and explanation tasks.
Definition
Codestral is Mistral AI's family of code-specialized large language models, first released in May 2024, trained on over 80 programming languages and optimized for code generation, completion, and explanation tasks.
Overview
Codestral was Mistral AI's first model dedicated specifically to coding tasks, released as a 22-billion-parameter model in May 2024. It was trained on a large corpus spanning more than 80 programming languages, including mainstream languages like Python, Java, JavaScript, C++, and Go as well as more specialized ones like Swift and Fortran, and was designed to handle both code completion (fill-in-the-middle style, useful for IDE autocomplete) and broader code generation and explanation tasks from natural language instructions. Mistral released Codestral under a license permitting free use for research and testing purposes, with separate commercial licensing terms required for production/commercial deployment — a different approach from the fully permissive Apache 2.0 license used for some other Mistral models. Codestral was made available via Mistral's API, through integrations with development tools and IDEs (including plugins for popular code editors), and via cloud platform partnerships. Mistral has continued to expand the Codestral line, including Codestral Mamba (a state-space-model-based coding model exploring an alternative to transformer architecture for faster inference on long code contexts) and subsequent Codestral version updates with improved benchmark performance. The model competes in the increasingly crowded code-LLM space alongside GitHub Copilot's underlying models, Meta's Code Llama, and other specialized coding models, as well as general-purpose frontier models like GPT-4 and Claude that also perform strongly on coding tasks; Codestral's differentiation is its focus specifically on being a fast, broad-language, IDE-integratable coding specialist.
Key Concepts
- Trained on over 80 programming languages
- Supports both fill-in-the-middle code completion and full code generation/explanation
- 22B parameter model at initial release, with subsequent version updates
- Codestral Mamba variant explores a state-space-model architecture for coding tasks
- Available via Mistral's API and IDE/editor plugin integrations
- Free for research/testing use with separate commercial licensing terms