Mistral AI Releases First Generative AI Model for Coding, 'Codestral,' Trained in 80+ Programming Languages



Mistral AI , an AI development company founded by former employees of Meta and Google DeepMind, has announced its first coding AI model, Codestral . Codestral is a generative AI model designed specifically for coding tasks, helping software developers design advanced AI applications.

Codestral: Hello, World! | Mistral AI | Frontier AI in your hands
https://mistral.ai/news/codestral/



Mistral releases Codestral, its first generative AI model for code | TechCrunch
https://techcrunch.com/2024/05/29/mistral-releases-its-first-generative-ai-model-for-code/

Codestral's AI models are trained on a diverse dataset of 80+ programming languages, including the most popular languages ​​such as Python, Java, C, C++, JavaScript, and Bash, and perform well in more specific programming languages ​​such as Swift and Fortran. The broad language base enables Codestral to assist developers in a variety of coding environments and projects.

Codestral can complete coding or partially complete code using an intermediate completion mechanism. By using Codestral, developers can not only improve their coding skills but also reduce the risk of errors and bugs. In addition, Codestral can answer questions about the code in English.



Codestral's context window (the number of tokens that an AI model can process at one time) is 32k (32,000 tokens), which is very large compared to existing coding AI. Below is a table comparing the scores of coding performance benchmarks with AI models from other companies. In the test,

HumanEval , MBPP , CruxEval-O , and RepoBench were used to measure coding performance in Python, Spider was used to measure coding performance in SQL, and HumanEval was used to compare average scores in multiple programming languages ​​(Python, C++, bash, Java, PHP, Typescript, and C#). FIM is a score that measures the intermediate completion function of each AI model. Although Codestral is not at the top of all scores, it recorded top-class scores in almost all benchmarks.



Although Mistral AI describes Codestral as open, it prohibits the use of Codestral or its deliverables in any commercial activities, and the license also explicitly prohibits 'any use of Codestral internally by employees in the context of the company's business activities.'

The reason for this is that 'Codestral may have been partially trained on copyrighted content,' TechCrunch pointed out. Mistral AI has not commented on this, but TechCrunch noted that ' Mistral AI's previous training datasets have included copyrighted data. '

TechCrunch also noted, 'Codestral, with its 22 billion parameters, requires a high-performance PC to run. Benchmarks show that it beats the competition, but not by a large margin.' 'Codestral is impractical for most developers, and its performance improvements are only incremental. But it's sure to spark discussion about relying on code-generation models as programming assistants.'

Related Posts:

in Software, Posted by logu_ii