Join our daily and weekly newsletters for the latest updates and exclusive content covering cutting-edge AI. Learn more
Mistral has updated its open source Codestral coding template – which is proving popular among coders – expanding competition for coding-focused templates aimed at developers.
In a blog postthe company said it upgraded the model with a more efficient architecture to create Codestral 25.01, a model that Mistral promises to be the “undisputed coding leader 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 actions and supports code correction, test generation, and middle-filling tasks. The company said this could be useful to companies with more data and model residence use cases.
Benchmark tests showed that Codestral 25.01 performed better in Python coding tests and scored 86.6% on 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 Mistral’s IDE plugin partners. Users can deploy Codestral 25.01 locally via Code Wizard Continue. They can also access the model API via the Mistral Platform and Google Vertex AI. The model is available in preview on Azure AI Foundry and will be available soon on Amazon Bedrock.
More and more coding models
Mistral released Codestral in May last year as its first code-focused model. The 22B parametric model could code in 80 different languages and outperformed other code-centric models. Since then, Mistral released Codestral-Mambaa code generation model built on the Mamba architecture that can generate longer code strings and handle more inputs.
And it looks like Codestral 25.01 is already generating a lot of interest. Just a few hours after Mistral’s announcement, the model is already climbing the rankings on Copilot Arena.
Writing code was one of the first features of core models, even for more general models like o3 from OpenAI And Claude d’Anthropice. However, over the past year, coding-specific models have improved and often outperform larger models.
In the past year alone, several coding-specific templates have been made available to developers. Alibaba released Qwen2.5-Encoder in November. China DeepSeek encoder became the first model to beat the GPT-4 Turbo in June. Microsoft too unveiled GRIN-MoEa mixture of experts (MOE) based model that can code and solve mathematical problems.
No one has resolved the eternal debate of choosing a general model that learns everything or a targeted model that only knows how to code. Some developers prefer the breadth of options they find in a template like Claude, but the proliferation of coding templates shows a demand for specificity. Since Codestral is trained in data coding, he will of course be better at coding tasks rather than email writing.
