French AI startup Mistral AI recently released Codestral, its first large language model (LLM) designed specifically for programming.
Codestral is an open AI model tailored for code generation tasks, providing instruction sharing and code completion API interfaces to help developers write and operate code more efficiently. Proficient in programming languages and English, Codestral enables software developers to create advanced AI applications.
The model has a parameter size of 22B, with 2.2 billion parameters, and adheres to the new Mistral AI Non-Production License, suitable for research and testing but not for commercial use.
Currently, the model can be downloaded from HuggingFace.
🔗Download Link🔗
https://huggingface.co/mistralai/Codestral-22B-v0.1
🔗Free Trial Link🔗
https://t.co/LsgC84GCYw
Guillaume Lample, co-founder and chief scientist of Mistral AI, stated that Codestral can be easily integrated into VS Code plugins.
Users comparing Codestral and GPT-4o noted that Codestral’s speed significantly surpasses that of GPT-4o.
Proficient in 80+ Programming Languages
Codestral has been trained on a vast dataset containing over 80 programming languages, including mainstream languages such as Python, Java, C, C++, JavaScript, and Bash. It also demonstrates outstanding performance in languages like Swift and Fortran.
This multilingual training foundation allows Codestral to adapt to various programming environments and projects, providing support for developers.
Codestral can not only write and test code but also use intermediate filling techniques to complete any part of the code, saving developers valuable time and effort. Using Codestral can also help improve developers’ programming skills and reduce the risk of errors and bugs.
A New Standard for Code Generation Performance
As a 22B parameter model, Codestral sets a new benchmark for performance and response time in code generation.
As shown in the data below, Codestral has a context window length of 32k, compared to CodeLlama (70B) at 4k, DeepSeek Coder (33B) at 16k, and Llama3 (70B) at 8k.
The results show that in remote evaluation benchmark testing for code generation performance on RepoBench, Codestral outperformed other competing models.
Mistral AI compared Codestral with existing models specifically designed for code, which typically require high hardware configurations.
For the Python programming language, Codestral’s code generation capability was evaluated using the HumanEval pass@1 and MBPP sanitised pass@1 benchmarks; additionally, CruxEval and RepoBench EM benchmarks were used for evaluation.
For the SQL language, researchers used the Spider benchmark to evaluate Codestral’s performance.
In other programming languages, Codestral was also evaluated in C++, Bash, Java, PHP, Typescript, and C#, and the average of these evaluations was calculated to comprehensively measure its performance across different programming languages.
In the FIM benchmark test, researchers evaluated Codestral’s ability to complete missing code in snippets.This experiment focused on Python, JavaScript, and Java.
The experimental results showed that the code snippets completed by Codestral could be run directly by users, indicating its efficiency and accuracy in code completion.
Reply “Programming” to the card to get1000+ Programming Learning Manuals
