Codestral 25.01 release notes / what's new

Mistral AI has announced the release of Codestral 25.01, a coding model that generates and completes code approximately two times faster than the original version. This update introduces a more efficient architecture and an improved tokenizer to establish a new state-of-the-art (SOTA) standard for fill-in-the-middle (FIM) use cases across the board.

Performance and Architecture Improvements

Codestral 25.01 utilizes a more efficient architecture and an improved tokenizer to double the speed of code generation and completion. The model is designed for low-latency, high-frequency use cases and supports critical developer tasks including fill-in-the-middle (FIM), code correction, and test generation. It is proficient in over 80 programming languages and supports a context length of 256k.

Technical Benchmarks

Codestral 25.01 outperforms leading sub-100B parameter coding models across several key benchmarks.

General Coding Proficiency

In HumanEval (average), Codestral 25.01 achieved 71.4%, surpassing the previous Codestral-2405 22B (65.6%), DeepSeek Coder V2 lite (65.9%), and Codellama 70B instruct (55.3%).

Fill-in-the-Middle (FIM) Capabilities

The model is described as SOTA for FIM use cases. In HumanEvalFIM (average) single line exact match, it scored 85.89%, beating the OpenAI FIM API (83.7%) and DeepSeek Chat API (84.63%).

For FIM pass@1, Codestral 25.01 achieved an average of 95.3%, compared to 91.8% for Codestral-2405 22B and 92.7% for the OpenAI FIM API.

Per-Language Performance

Codestral 25.01 leads in most tested languages on HumanEval, including Python (86.6%), C++ (78.9%), Javascript (82.6%), Bash (43.0%), and Typescript (82.4%). The only exception in the provided data is Java, where Codestral-2405 22B scored higher at 78.5% compared to 25.01's 72.8%.

Deployment and Availability

Codestral 25.01 is available through the following channels:

  • IDE Integration: Available via IDE and IDE plugin partners, including as a recommended autocomplete model in Continue for VS Code and JetBrains.
  • API Access: Accessible via codestral-latest on la Plateforme, Google Cloud's Vertex AI, and in private preview on Azure AI Foundry, with Amazon Bedrock support coming soon.
  • Enterprise Deployment: The model can be deployed locally within a company's premises or Virtual Private Cloud (VPC) for organizations requiring data and model residency.

Industry Perspective

Ty Dunn, co-founder of Continue, emphasizes the importance of FIM capabilities for AI code assistants:

"For AI code assistants, code completion constitutes a large portion of the work, which requires models that are great at fill-in-the-middle (FIM). Codestral 25.01 marks a significant advancement in this area. Mistral AI’s new model is capable of providing more precise suggestions, much faster — a critical component of accurate, efficient software development."

Codestral 25.01 debuted at #1 on the LMsys copilot arena leaderboard.

Sources

Related

  • Dispatch
    Codestral Release Notes

    Mistral AI has released Codestral, a 22B open-weight generative AI model specifically designed for code generation across 80+ programming languages.

  • Dispatch
    Mistral AI Codestral 25.08 and Enterprise Coding Stack Release

    Mistral AI has released Codestral 25.08 and a comprehensive enterprise coding stack featuring integrated completion, semantic search, and agentic workflows with flexible deployment options.

  • Dispatch
    Mistral AI Devstral Models Release

    Mistral AI has released Devstral Medium and Devstral Small 1.1, providing state-of-the-art agentic coding capabilities with a focus on generalization across different prompts and agentic scaffolds.

  • Dispatch
    Mistral Code Release Notes

    Mistral AI has launched Mistral Code, a vertically integrated AI coding assistant for enterprises that combines frontier models with an in-IDE assistant and flexible deployment options.

  • Dispatch
    Codestral Embed model release

    Mistral AI released Codestral Embed, a code‑specialized embedding model that outperforms existing code embedders on retrieval benchmarks and supports flexible dimensions and precisions.