Codestral

Generates code from natural language instructions using a 22B parameter decoder-only transformer trained on 80+ programming languages. Processes up to 32K tokens of context (approximately 24K tokens of code + instructions), enabling multi-file code generation and understanding of large codebases within a single request. Implements standard instruction-following fine-tuning patterns built into the base model training rather than separate RLHF stages.

Implements fill-in-the-middle (FIM) mechanism enabling IDE plugins to request code completion at arbitrary positions within a file by providing prefix and suffix context. The model processes both left and right context to predict the missing middle section, supporting real-time IDE workflows where users type in the middle of incomplete code. Requires specific prompt formatting (details not disclosed) and routes through dedicated codestral.mistral.ai endpoint optimized for low-latency IDE requests.

Codestral weights distributed under Mistral AI Non-Production License restricting use to research, testing, and evaluation. Commercial use requires explicit commercial license agreement from Mistral AI with terms and pricing determined on case-by-case basis. Enables free evaluation and research while protecting Mistral's commercial interests through licensing restrictions.

Generates SQL queries from natural language descriptions or existing database schemas. Evaluated on Spider benchmark (complex SQL generation from text) but specific scores not disclosed. Supports SQL generation for various databases and query types as part of 80+ language support.

Codestral FIM capability evaluated against DeepSeek Coder 33B on HumanEval pass@1 metrics across Python, JavaScript, and Java, demonstrating competitive FIM performance despite smaller parameter count (22B vs 33B). Evaluation highlights efficiency advantage of smaller model with comparable FIM quality.

Trained on diverse dataset spanning 80+ programming languages including Python, JavaScript, TypeScript, Java, C++, C, Rust, Go, PHP, C#, Swift, Bash, SQL, Fortran and others. Model learns language-specific syntax, idioms, and patterns through unified transformer architecture rather than language-specific models. Supports code generation, completion, and instruction-following in any of the 80+ languages with single model inference.

Generates unit tests, integration tests, and validation code from function signatures, docstrings, and existing code. Evaluated on MBPP (Mostly Basic Python Programming) benchmark for test generation capability. Synthesizes test cases that cover edge cases, error conditions, and normal operation paths based on code context and instruction prompts.

Leverages 32K token context window to maintain understanding of large code repositories and multi-file dependencies. Evaluated on RepoBench benchmark for repository-level code completion where model must understand cross-file references, imports, and function definitions across multiple files. Outperforms competitors on RepoBench according to source material, enabling code generation that respects existing codebase patterns and dependencies.

Provides two distinct API endpoints for different use cases: (1) codestral.mistral.ai — dedicated endpoint for IDE plugins with free beta access (8 weeks), personal API key management, and optimized latency for real-time completion; (2) api.mistral.ai — standard endpoint with token-based billing, organization-level rate limits, and support for batch queries and third-party applications. Both endpoints support streaming responses for real-time output display.

Codestral weights available for download via HuggingFace, enabling self-hosted inference on local hardware or private infrastructure. Model distributed in open-weight format (specific serialization format not disclosed — likely safetensors or GGUF) under Mistral AI Non-Production License. Supports local deployment without API calls, enabling offline code generation, private data handling, and custom fine-tuning.

Evaluated on multiple code generation benchmarks: HumanEval (Python function generation), MBPP (Mostly Basic Python Programming for test generation), CruxEval (Python output prediction), RepoBench (repository-level code completion), Spider (SQL generation), and multi-language HumanEval variants (C++, Bash, Java, PHP, TypeScript, C#). Provides comparative performance claims across diverse code generation tasks without disclosing absolute scores.

Accepts natural language instructions and generates corresponding code without requiring specific prompt templates or few-shot examples. Instruction-following capability built into base model training rather than requiring separate fine-tuning. Supports diverse instruction types: function generation from descriptions, code refactoring requests, documentation generation, and code explanation tasks.

Both API endpoints (codestral.mistral.ai and api.mistral.ai) support streaming responses where generated code is returned as a stream of tokens rather than waiting for full completion. Enables real-time display of generated code in IDEs and web interfaces as tokens are produced, improving perceived latency and user experience. Streaming tokens can be displayed incrementally without waiting for full response.