Mistral unveils AI code-generation model Codestral

The model has already undergone testing by JetBrains, SourceGraph, LlamaIndex and others

Mistral, the $6 billion French AI start-up backed by Microsoft, has entered the coding assistant market with the launch of Codestral, its first generative AI tool specifically designed for developers

Similar to existing AI tools, Codestral helps developers write and interact with code across a range of over 80 programming languages, including Python, Java, C++ and JavaScript.

Codestral is a 22 billion parameter model, with functionality like completing code functions, writing tests and "filling in" partial code. It can also answer questions about a codebase in plain English.

Mistral claims that Codestral outperforms previous coding-focused models, including CodeLlama 70B and Deepseek Coder 33B. The company believes the new tool will accelerate workflows and saving developers significant time and effort.

While new to the public, the model has already undergone testing by industry partners like JetBrains, SourceGraph and LlamaIndex.

"Despite its relatively compact size, it delivers results on par with much larger models we offer to customers," said Meital Zilberstein, R&D lead at AI coding assistant firm Tabnine.

While Mistral describes Codestral as an "open-weight" generative model, it's not strictly open source in the traditional sense.

The model falls under the newly introduced Mistral AI Non-Production Licence, permitting its use for "research and testing purposes." This essentially limits its application for purely commercial endeavours.

Mistral offers Codestral on Hugging Face under a non-production licence.

Additionally, two API endpoints are available:

• codestral.mistral.ai: Designed for developers to use Codestral's Instruct or Fill-in-the-Middle functions within their IDEs. Free to use during an eight-week beta period. Comes with a personal API key.
• api.mistral.ai: This standard endpoint caters to broader research, batch queries, and third-party application development, with billing based on token usage.

Codestral's launch positions Mistral as a serious contender in the coding assistant space, challenging established players like GitHub, AWS and Meta.

However, it also fuels the ongoing debate about relying on AI-generated code as a programming assistant.

A June 2023 Stack Overflow poll revealed that 44% of developers already use AI tools, with 26% planning to do so soon. However, these tools have obvious flaws.

An analysis by GitClear of over 150 million code lines found generative AI tools contributing to more incorrect code being pushed to repositories. Other studies have highlighted the issue of accuracy, with over half of ChatGPT's responses to programming questions being wrong.

Security researchers have also expressed concerns about these tools' potential to amplify existing bugs and vulnerabilities.