Codestral by Mistral

freemium

Codestral is Mistral AI's 22B open-weight model for code generation, supporting 80+ programming languages with fill-in-the-middle, code completion, and test writing capabilities.

Coding & Development

Foundation Models

LLM Developer Tools

About

Codestral is Mistral AI's flagship open-weight generative model purpose-built for code generation. Released in May 2024, it is a 22B parameter model trained on a diverse dataset spanning 80+ programming languages — from mainstream options like Python, Java, C++, JavaScript, and Bash to more specialized ones like Swift and Fortran. This broad language coverage makes it a versatile assistant across virtually any coding environment. Codestral excels at completing coding functions, writing unit tests, and finishing partial code using a fill-in-the-middle (FIM) mechanism. Its 32k context window surpasses competitors operating at 4k–16k, giving it a significant edge in repository-level code understanding and long-range completions. On benchmarks like HumanEval, MBPP, CruxEval, RepoBench, and Spider (SQL), Codestral outperforms existing code-specific models relative to its hardware footprint. Developers can access Codestral in multiple ways: download the open-weight model from HuggingFace for research and testing under the Mistral AI Non-Production License, use the dedicated `codestral.mistral.ai` endpoint optimized for IDE plugins and personal API key integrations, or query it through the standard `api.mistral.ai` endpoint billed per token for production and batch workloads. Commercial licenses are available on request. Codestral is ideal for individual developers building AI-powered coding tools, IDE plugin creators, and organizations looking to integrate state-of-the-art code intelligence into their software workflows.

Key Features

80+ Programming Language Support: Trained on a diverse dataset covering over 80 languages including Python, Java, C++, JavaScript, Bash, Swift, Fortran, SQL, and more.
Fill-in-the-Middle (FIM) Mechanism: Completes partial code snippets by understanding surrounding context — ideal for in-editor autocomplete and inline suggestions.
32k Token Context Window: A large 32k context window enables repository-level code completion and outperforms competitors on long-range benchmarks like RepoBench.
Dedicated IDE-Optimized Endpoint: The codestral.mistral.ai endpoint is designed for IDE plugins and tools where developers bring their own API keys, free during beta.
Code Completion & Test Writing: Automates repetitive development tasks such as writing unit tests, completing functions, and finishing partial implementations.

Use Cases

Integrating AI-powered autocomplete and inline code suggestions into IDEs like VS Code or JetBrains via the Codestral API endpoint.
Automatically generating unit tests for existing functions to improve code coverage without manual effort.
Completing partial or boilerplate code snippets using fill-in-the-middle to accelerate development workflows.
Building SQL queries and interacting with databases using natural language through Codestral's strong SQL benchmark performance.
Developing custom AI coding assistants and developer tools that expose Codestral's capabilities to end users via the production API.

Pros

Industry-Leading Context Window: With 32k tokens, Codestral handles complex, multi-file repository contexts that smaller-window models cannot, enabling smarter completions.
Open-Weight & Downloadable: Available on HuggingFace for research and testing, giving developers full model access and the ability to self-host or fine-tune.
Strong Benchmark Performance: Outperforms many larger code-specific models on HumanEval, MBPP, CruxEval, RepoBench, and Spider despite its efficient 22B size.
Flexible Access Options: Multiple deployment paths — open-weight download, dedicated personal endpoint, or production API — suit a range of use cases and budgets.

Cons

Non-Production License for Free Use: The open-weight model is free only for research and testing; commercial use requires a separate paid license from Mistral AI.
Hardware Demands for Self-Hosting: At 22B parameters, running Codestral locally requires substantial GPU memory, making self-hosting impractical for many individual developers.
Beta Endpoint Access Gated by Waitlist: The free dedicated endpoint (codestral.mistral.ai) was initially gated behind a waitlist, which may slow onboarding for some developers.

Frequently Asked Questions

Codestral is Mistral AI's first dedicated code generation model — an open-weight 22B parameter model trained on 80+ programming languages, designed to help developers write, complete, and test code more efficiently.

Codestral supports 80+ programming languages, including Python, Java, C, C++, JavaScript, Bash, PHP, TypeScript, C#, Swift, Fortran, and SQL, among many others.

Codestral is free to download and use for research and testing under the Mistral AI Non-Production License. A dedicated API endpoint was also offered free during a beta period. Commercial use requires a paid license.

FIM is a technique where the model completes code given both the preceding and following context — essential for in-editor autocomplete. Codestral's FIM performance is benchmarked against DeepSeek Coder 33B across Python, JavaScript, and Java.

You can use the dedicated codestral.mistral.ai endpoint with a personal API key for IDE plugin integrations, or the standard api.mistral.ai endpoint for production applications billed per token. The model can also be downloaded from HuggingFace for local use.