New Codestral 25.01 sets FIM performance benchmark in AI coding

Mistral AI has announced the release of Codestral 25.01, an upgraded version of its state-of-the-art coding model designed to enhance software development workflows. This release builds on the success of the original Codestral model, which has been widely adopted by developers for tasks such as code generation, fill-in-the-middle (FIM), code correction, and test generation.

Codestral 25.01

Codestral 25.01 introduces several advancements over its predecessor:

1. Improved Architecture: The model is now approximately twice as fast in generating and completing code, thanks to architectural optimizations and a more efficient tokenizer.
2. Expanded Context Length: With a context length of 256k tokens, it supports larger codebases and more complex tasks.
3. Performance Benchmarks: Codestral 25.01 outperforms leading sub-100B parameter coding models in FIM tasks across multiple programming languages. For instance:
   1. It achieves an average FIM pass@1 score of 95.3%, surpassing competitors like OpenAI's GPT-3.5 Turbo API and DeepSeek Coder models.
   2. On HumanEval benchmarks, it scores 86.6% for Python and maintains strong performance across other languages such as JavaScript (82.6%) and Java (72.8%).

The model is tailored for high-frequency, low-latency use cases, making it ideal for developers seeking faster and more accurate code assistance.

The Company Behind It: Mistral AI

Mistral AI has positioned itself as a leader in AI-driven coding solutions. Known for its lightweight and efficient models, the company focuses on improving developer productivity through cutting-edge tools like Codestral.

Availability

Codestral 25.01 is available starting today:

1. Developers can access it through IDE plugins for platforms like Visual Studio Code and JetBrains.
2. For enterprise users requiring local deployment or data residency, the model is offered via Continue’s platform.
3. The API is accessible on Google Cloud’s Vertex AI, with private previews on Azure AI Foundry and plans to launch on Amazon Bedrock.

Why This Matters

This release underscores Mistral AI's commitment to advancing coding models that cater to real-world developer needs. By prioritizing FIM capabilities and efficiency, Codestral 25.01 addresses critical challenges in software development, such as reducing coding errors and accelerating project timelines.

With its debut at #1 on the LMsys copilot arena leaderboard, Codestral 25.01 sets a new benchmark for coding assistants in its category. Developers worldwide are encouraged to explore its capabilities through free demos or integrate it into their workflows via APIs or IDE plugins.

Source