Codestral 25.01

Released January 13, 2025, Codestral 25.01 is Mistral's updated code generation model built for high-frequency, low-latency coding assistance. Its headline strength is fill-in-the-middle (FIM) completion: 95.3% average pass@1 across languages, topping sub-100B code models and beating OpenAI's FIM API. A new architecture and tokenizer doubled generation speed versus the original Codestral, the context window grew 8x to 256K tokens, and it covers 80+ programming languages with 86.6% on Python HumanEval. It topped LMsys' Copilot Arena at launch and ships free inside the Continue VS Code/JetBrains extension.\n\nLimitations: it is released under Mistral's Non-Production License (MNPL) — weights are downloadable but restricted to research, testing, and evaluation, not production deployment. It is a completion specialist, not a chat agent; complex reasoning and agentic planning belong to larger models. Fill-in-the-middle quality still trails dedicated agentic coding models on multi-file refactors.

Mistral-reported: 95.3% FIM pass@1, 86.6% HumanEval Python