Mistral's new Codestral code completion module races up third-party charts

Mistral has updated its Codestral open source coding module – popular among coders – widening the competition for coding-focused modules aimed at developers.

In a blog postthe company said it has updated the model with a more efficient architecture to create Codestral 25.01, a model that Mistral promises to be the “clear leader for coding in its weight class” and twice as fast as the previous version .

Like the original Codestral, Codestral 25.01 is optimized for low-latency, high-frequency tasks and supports code debugging, test generation and central fill functions. The company said it could be helpful for enterprises with more data and model residential use cases.

Tests showed that Codestral 25.01 performed better in coding tests in Python and scored 86.6% in the HumanEval test. It beat the previous version of Codestral, Codellama 70B Instruct and DeepSeek Coder 33B instruct.

This version of Codestral will be available to developers who are part of the Mistral IDE plugin partners. Users can use Codestral 25.01 locally via the code assistant Continue on. They can also access the API of the model through Plateforme Mistral and Google Vertex AI. The model is available in preview on Azure AI Foundry and will be on Amazon Bedrock soon.

More and more coding modules

Mistral released Codestral last May as the first code-oriented model. The 22B parameter model could code in 80 different languages ​​and outperformed other kernel code models. Since then, Mistral Codestral-Mamba releasea code generation module built on top of the Mamba architecture that can generate longer code strings and handle more input.

And, there seems to be a lot of interest already in Codestral 25.01. Just a few hours after Mistral published the news, the model is already building the leaderboards on Copilot Arena.

Writing code was one of the earliest features of base models, even for general purpose models like OpenAI's o3 and Claude Anthropic. However, in the past year, special code models have evolved, and often perform better than larger models.

In the past year alone, developers have been given a number of unique coding modules. Alibaba released Qwen2.5-Coder in November. China Code DeepSeek to be the first model to beat the GPT-4 Turbo in June. Microsoft too published GRIN-MoEa combination of model-based experts (MOE) that can encode and solve mathematical problems.

No one has solved the eternal debate of choosing a general purpose model that learns everything or a focused model that only knows how to code. Some developers prefer the range of options they find in a model like Claude, but the proliferation of coding models shows a demand for specificity. Since Codestral is trained on data coding, he will, of course, be better at coding tasks than writing emails.