Ad
Skip to content
Subscribe Now
Plus AI in practice
Copy the url to clipboard Share this article Go to comment section

Mistral's new code model features a 32K context window for unmatched long code support

Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
May 29, 2024
Image description
Mistral

French AI company Mistral has launched Codestral, a new coding model that delivers high coding performance with less computational overhead than existing models.

According to Mistral, Codestral handles more than 80 programming languages, including common ones such as Python, Java, C, C++, JavaScript, and Bash, as well as more specialized ones such as Swift and Fortran. Features include code completion, test writing, and filling in incomplete code with a fill-in-the-middle mechanism.

As a 22-billion-parameter model, Codestral sets a new standard for the performance/latency ratio of code generation compared to existing models, Mistral claims. With its larger context window of 32,000 tokens, Codestral outperforms all other models in RepoBench, a benchmark for longer code generation.

Codestral is supposed to be especially good at tasks with long code thanks to its 32K context window. | Image: Mistral

Mistral compared the performance of Codestral in various benchmarks for Python, SQL and other languages with competing models that have higher hardware needs. Codestral consistently performed better, e.g. in completing code repositories over long distances or predicting Python output.

In the prestigious HumanEval code benchmark, Codestral slightly outperforms the much larger Llama 3 70B. | Image: Mistral

Codestral is licensed as an open-weight model under the new Mistral AI Non-Production License, which allows it to be used for research and testing purposes. It can be downloaded from HuggingFace.

Mistral also provides two API endpoints for Codestral: codestral.mistral.ai for integration into IDEs, where developers bring their own API keys, and api.mistral.ai for research, batch queries or application development, where results are shown directly to users. The former is free for eight weeks, while the latter is charged per token.

Some early feedback from developers and researchers supports Codestral's good performance. Despite its relatively small size, it delivers results similar to larger models.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Read on for the full picture.
Subscribe for hype-free coverage.

  • Full access to every article on THE DECODER
  • No ads
  • Join the comments and community discussions
  • A weekly AI news recap via mail
  • 6x/year: "AI Radar" — deep dives on the AI topics that matter most
  • Daily AI news, always up to date
  • Our full ten-year archive
  • Covered by a team with 10+ years in AI
Subscribe to The Decoder