Mistral Corporation recently released a major update to its open source code generation model Codestral - Codestral25.01. This version is architecturally optimized, twice as fast as the previous generation, and performs well in various benchmark tests, especially in the Python coding test. Codestral25.01 supports multiple functions such as code correction, test generation and code filling, aiming to provide developers with a more efficient code generation experience. It has quickly gained market attention and topped the Copilot Arena rankings. Mistral provides multiple ways to access Codestral25.01, including IDE plug-ins, Mistral platform, Google Vertex AI and Azure AI Foundry preview versions, and will also be available on Amazon Bedrock in the future.
According to Mistral's blog post, Codestral 25.01's architecture has been optimized and promises to be the "absolute leader" among similar models, with code generation twice as fast as previous versions. This new version still maintains the advantages of the original model, focusing on low-latency and high-frequency operations, supporting tasks such as code correction, test generation, and code filling. Mistral says this is especially important for enterprises with large amounts of data and model-resident use cases.
In various benchmark tests, Codestral25.01 performed well in the Python coding test, achieving a high score of 86.6% in the HumanEval test, surpassing previous versions of Codestral, Codellama70B Instruct and DeepSeek Coder33B Instruct.
Developers can use Codestral25.01 through Mistral's IDE plug-in partners. In addition, users can access the model's API through Mistral's platform and Google Vertex AI. The model is currently available in preview on Azure AI Foundry and will be available on Amazon Bedrock.
Since the first release of Codestral in May 2023, Mistral has continuously promoted the upgrade and innovation of its products. The previously introduced Codestral-Mamba model is based on the Mamba architecture and can generate longer code strings and handle more inputs. It is worth noting that Codestral25.01 has quickly climbed to the C opilot Arena rankings within a few hours of Mistral’s announcement, showing the market’s strong interest in this new model.
Writing code is one of the early features of the basic model. Although it is also used in general models such as OpenAI’s o3 and Anthropic’s Claude, in the past year, models focused on programming have made significant progress, often surpassing some large-scale models. Universal model. Recently, Alibaba, DeepSeek Coder and Microsoft have also released new programming models, and competition has become increasingly fierce.
Among many developers, there is still debate over whether to choose a general-purpose model or a programming-focused model. Some developers prefer to use general-purpose models like Claude, while demand for programming tasks drives the emergence of specialized models. Because Codestral is trained specifically on coded data, it naturally performs better on programming tasks.
Official blog: https://mistral.ai/news/codestral-2501/
Highlight:
Mistral launches Codestral 25.01, which increases code generation speed to twice that of the previous version.
The model performed well in the Python coding test, obtaining a HumanEval test score of 86.6%.
Codestral25.01 quickly rose to the top of the C opilot Arena and received widespread attention from developers.
The release of Codestral25.01 marks a further intensification of competition in the field of code generation models. Its excellent performance and convenient access make it a new tool worthy of developers' attention. It is worth looking forward to how Mistral will further improve Codestral in the future.