Mistral AI, the French company behind the cutting-edge Mistral and Mixtral language models, has announced Codestral, its first LLM designed specifically for coding help. The company said that its newest LLM is fluent in over 80 programming languages including Python, Java, C, C++, JavaScript, Bash, Swift, and even Fortran which came out in 1957.
Similar to other coding LLMs, Codestral can help developers complete functions, write tests, and complete any partial code using a fill-in-the-middle mechanism. According to the company, developers and the companies they work for can benefit from using Codestral because it will reduce the risk of errors and bugs in their code.
In the announcement, Mistral AI compared Codestral 22B against CodeLlama 70B, DeepSeek Coder 33B, and Llama 3 70B. Despite fewer parameters used by Codestral, it has a larger 32k context window and generally performs better at Python, SQL, and other languages according to several benchmarks that were carried out. Llama 3 70B was able to outperform Codestral 22B in the SQL spider benchmark scoring 67.1% compared to 63.5% and DeepSeek Code 3B was able to beat Codestral 22B 80.2% to 78.2% in the MBPP Python benchmark but in everything else, Codestral was the one to beat despite its fewer parameters.
With fewer parameters, it means that the computation power needed to run the model isn’t as high. This will benefit anyone using this LLM on platforms like Poe.com where you can send more queries to smaller models.
As a bit of background, Mistral AI is one of the lesser-known players in the generative AI space but is still well-known in the space. Back in February, Microsoft struck a deal with Mistral AI so that it could use its models on its Azure platform. The deal also means that Azure AI Studio customers could use Mistral Large, the company’s most powerful model.
There were reports that Microsoft invested capital in Mistral AI but the amount wasn’t disclosed. Mistral AI was set up in 2023 by AI engineers formerly from Meta and Google. In the same year, it reportedly raised around $415 million to help with its work.
To get started with Codestral, head over to Le Chat and under the prompt box press the model drop down and select Codestral.
Source: Mistral AI