Join our daily and weekly newsletters for the latest industry-leading AI updates and exclusive content. learn more
Mistral has updated its open source coding model Codestral. Codestral is popular among programmers and is expanding the competition for coding-focused models aimed at developers.
in a blog postThe company said it upgraded the model with a more efficient architecture to create Codestral 25.01. The model promises to make Mistral “the clear leader in coding in its class” and to be twice as fast as previous versions.
Like the original Codestral, Codestral 25.01 is optimized for low-latency, high-frequency actions, supporting code modification, test generation, and intermediate completion tasks. The company said it could be useful for companies with more data and model-resident use cases.
Benchmark tests showed that Codestral 25.01 performed better on coding tests in Python, scoring 86.6% on the HumanEval test. This exceeds Codellama 70B and DeepSeek Coder 33B instructions, previous versions of Codestral.
This version of Codestral is available to developers belonging to Mistral’s IDE plugin partners. Users can deploy Codestral 25.01 locally through Code Assistant. Continue. You can also access the model API through Mistral’s Plateforme and Google Vertex AI. This model is available in preview on Azure AI Foundry and will soon be available on Amazon Bedrock.
Coding models are increasing rapidly
Last May, Mistral released Codestral, its first code-focused model. The 22B parameter model could be coded in 80 different languages and performed better than other code-centric models. Since then, Mistral has released Codestral-Mamba, a code generation model built on the Mamba architecture that can generate longer code strings and handle more input.
And Codestral 25.01 already seems to be generating a lot of interest. Just a few hours after the Mistral was announced, the model is already climbing up the CoPilot Arena leaderboard.
Writing code was one of the early features of foundational models, even more general models like OpenAI’s o3 and Anthropic’s Claude. However, over the past year, coding-specific models have improved and often perform better than larger models.
In the past year alone, there have been several coding-specific models available to developers. Alibaba released Qwen2.5-Coder in November. China’s DeepSeek Coder became the first model to defeat GPT-4 Turbo in June. Microsoft also announced GRIN-MoE, a Mixed Expertise (MOE)-based model that allows you to code and solve math problems.
No one has resolved the eternal debate of choosing a general-purpose model that learns everything or a focused model that only knows how to code. While some developers prefer the wide range of options found in models like Claude, the proliferation of coding models demands specificity. Since Codestral is trained in data coding, it is naturally better at coding tasks than writing emails.
