AI BriefingMAI-Code-1.1-Flash: Better, Faster, and a Quarter of the Cost
Key point
Microsoft AI introduced MAI-Code-1.1-Flash to GitHub Copilot, improving performance and speed while reducing costs to a quarter.
Details
MAI-Code-1.1-Flash generates higher-quality code compared to the model unveiled at Microsoft Build in June, with a 25% improvement in token efficiency and costs reduced to one-quarter of the previous level. This efficient coding model is currently deployed in the GitHub Copilot production environment.
By incorporating developer feedback and focusing on CLI tasks and .NET performance, the following performance improvements were achieved:
- Terminal-Bench 2.1 (GitHub Copilot CLI): 22% performance improvement
- .NET tasks: 15% performance improvement
Positive changes were also confirmed in actual production metrics. Code survival rates increased by 4%, and user return rates rose by 9%.
Additionally, model efficiency was maximized, resulting in 25% faster token streaming speeds in GitHub Copilot and a 25% reduction in token usage required to complete tasks. This means more work can be done with faster responses and lower costs.
These achievements were realized by optimizing real-world use cases through hundreds of thousands of Reinforcement Learning environments within GitHub Copilot.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.