DMR News

Advancing Digital Conversations

Mistral Debuts Codestral, Its First Code-Focused AI Model

ByHilary Ong

Jun 2, 2024

Mistral Debuts Codestral, Its First Code-Focused AI Model

Mistral, the Paris-based AI startup, has launched Codestral, its first code-centric large language model (LLM). Codestral, a 22B parameter model now available under a non-commercial license, caters to a variety of coding tasks, from generation to completion. This model, which supports over 80 programming languages, is designed to help developers create more sophisticated AI applications efficiently.

As a newly emerged leader in the global AI sector, following its record seed funding in Europe, Mistral has quickly made its mark with Codestral by claiming superior performance over existing models like CodeLlama 70B and Deepseek Coder 33B. The model is already in use by notable industry partners such as JetBrains, SourceGraph, and LlamaIndex, emphasizing its early adoption and potential impact on software development.

Codestral performance on HumanEval across different programming languages
Credits: Mistral

Features and Performance of Codestral

Codestral stands out with its ability to handle a massive context length of 32K, allowing for versatile coding interactions. Trained on a dataset encompassing a wide array of programming languages—from mainstream ones like SQL, Python, Java, C, and C++ to niche ones like Swift and Fortran—Codestral is equipped to handle diverse coding scenarios. This includes generating code from scratch, completing functions, writing tests, and filling in missing code segments.

The model’s proficiency is evident in its benchmarking scores, where it has outperformed leading competitors across various tests. On RepoBench, Codestral led with a 34% accuracy score for Python code completion. In other evaluations like HumanEval and CruxEval, it scored 81.1% and 51.3%, respectively, showcasing its capability in generating and predicting Python code. Although it faced some challenges with C++, C, and Typescript, its average performance score across all tests was a leading 61.5%.

Where to Access Codestral?

To facilitate easy access and integration for developers, Codestral is available on Hugging Face with a non-production license, ideal for non-commercial use, testing, and research. Mistral provides two API endpoints: one for specific instructive and fill-in-the-middle coding tasks within an IDE, and another for broader research and application development purposes

Several development tools and platforms have begun incorporating Codestral, with LangChain’s CEO Harrison Chase noting the model’s effectiveness in self-corrective code generation and its robust performance out-of-the-box.

The introduction of Codestral adds a significant player to the increasingly competitive AI-driven coding tool market. It now stands alongside established offerings like OpenAI’s Codex, Amazon’s CodeWhisper, and other emerging tools such as StarCoder2 and Codenium. This diversity in tools provides developers with numerous options to streamline their coding processes, reduce errors, and accelerate development timelines.


Related News:


Featured Image courtesy of Mohamed Ahmed Soliman/Dreamstime.com

Hilary Ong

Hello, from one tech geek to another. Not your beloved TechCrunch writer, but a writer with an avid interest in the fast-paced tech scenes and all the latest tech mojo. I bring with me a unique take towards tech with a honed applied psychology perspective to make tech news digestible. In other words, I deliver tech news that is easy to read.

Leave a Reply

Your email address will not be published. Required fields are marked *