French AI Startups Mistral AI Today we announced Codestral, the company’s first large language model (LLM) designed specifically for programming.
Codestral is proficient in more than 80 programming languages, including Python, Java, C, C++, JavaScript, Bash, Swift, and even Fortran, which was invented in 1957.
Similar to other programming LLMs, Codestral can help developers complete various functions, including completing functions, writing tests, filling in code, etc., reducing the risk of errors and vulnerabilities in the code.
Mistral AI also used a chart to compare Codestral 22B with similar large models such as CodeLlama 70B, DeepSeek Coder 33B and Llama 3 70B, indicating that although Codestral has only 22 billion parameters, its context length reaches 32,000 tokens. According to several benchmark tests conducted, it generally performs better in Python, SQL and other languages.
In the SQL spider benchmark, Llama 3 70B scored 67.1% and Codestral 22B scored 63.5%; in the MBPP Python benchmark, DeepSeek Code 33B scored 80.2% and Codestral 22B scored 78.2%.