Mistral Launches Codestral: The First Generative AI Model for Coding Solutions

Mistral, the French AI startup valued at $6 billion and backed by Microsoft, has introduced its first generative AI model for coding, named Codestral. This innovative model is designed to assist developers in writing and interacting with code, being trained on over 80 programming languages, including Python, Java, C++, and JavaScript, as mentioned in Mistral's blog post. Codestral can effectively complete coding functions, write tests, fill in incomplete code segments, and answer questions regarding codebases in English.

Mistral characterizes Codestral as “open,” but the terms of use raise questions. The licensing agreement prohibits any commercial use of Codestral and its outputs. Although there is an allowance for “development,” it comes with significant restrictions, explicitly banning “any internal usage by employees related to the company’s business activities.” This cautious approach may stem from the possibility that Codestral was partially trained on copyrighted content. While Mistral didn’t address this issue directly, previous evidence suggested that their training datasets included copyrighted material.

However, Codestral may not prove practical for many developers. With 22 billion parameters, the model demands a powerful PC to operate efficiently. Parameters serve as a measure of an AI model's proficiency in tasks like text analysis and generation. While it may outperform competitors based on some benchmarks—though these can be unreliable—it does not significantly surpass them.

Although Codestral might present challenges in real-world application and show only modest performance improvements, it is sure to ignite discussions about the efficacy of code-generating models as programming tools. Many developers are embracing generative AI tools for specific coding tasks. A June 2023 Stack Overflow poll indicated that 44% of developers currently use AI tools in their development processes, with an additional 26% planning to adopt them soon. Nonetheless, these tools are not without their issues.

An analysis from GitClear, examining over 150 million lines of code committed to project repositories in recent years, found that generative AI development tools often contribute to an increase in erroneous code being integrated into codebases. Additionally, security experts warn that these tools can exacerbate existing bugs and vulnerabilities in software projects; research from Purdue reveals that more than half of the programming answers provided by OpenAI’s ChatGPT are incorrect.

Despite these challenges, companies like Mistral continue to seek monetization and recognition for their models. Recently, Mistral launched a hosted version of Codestral on its Le Chat conversational AI platform, along with a paid API. The company has also integrated Codestral into various app frameworks and development environments, such as LlamaIndex, LangChain, Continue.dev, and Tabnine.