Mistral’s big language model focused on coding tasks: Codestral

European-based artificial intelligence company Mistralthe first code-centric large language model (LLM) CodestralHe announced. Codestral focuses on coding tasks from code generation to code completion. With 22 billion parameters, Codestral appears as a predominantly open generative artificial intelligence model.

Mistral made Codestral available today under a non-commercial license. According to the information shared by Mistral, the model has mastered more than 80 programming languages. Among these languages SQL, Python, Java, C ve C++ as well as popular languages ​​such as Swift ve Fortran There are more specific languages ​​such as:

Codestral 22B, Context window of 32 thousand tokens owner. Model developers can use both in various coding environments and in their projects. to write code ve interact with code provides. Tasks the model can perform include creating code from scratch, completing code writing functions, writing tests, and completing any partial code using the middle fill mechanism. Developers can benefit from Codestral to level up their projects and reduce the risk of errors and bugs.

However, according to the information shared, Codestral outperforms previous models designed for coding tasks, such as Meta’s CodeLlama 70B and DeepSeek AI’s Deepseek Coder 33B. of the model 34 percent accuracy on RepoBench It seems that it performs better than CodeLlama 70B, Deepseek Coder 33B, and Llama 3 70B with its score. Likewise, in HumanEval to evaluate Python code generation and CruxEval to test Python output prediction, the model is respectively 81.1 percent ve 51.3 percent surpassed its rivals with points. Moreover, Codestral 22B, Bash, Java ve PHP for HumanEvalIt performed better than the models in .

Developers can try Codestral 22B on Hugging Face. Moreover codestral.mistral.ai ve api.mistral.ai The model can also be accessed via . In addition, we should also point out that in Le Chat, Mistral’s free conversation interface, the model can chat with a version of the model that has been specially trained with various instructions. When we look at the industry partners using the model, we see names such as SourceGraph, LlamaIndex, LangChain, Continue.dev, Tabnine and JetBrains.