News from the IT community on October 8th: At an event held in San Francisco, USA, on October 7th at 10 a.m. Pacific Time (October 8th at 1 a.m. Beijing Time), Microsoft announced that it plans to introduce the MAI Code 1.1 Flash model (which contains 130 billion parameters) to Windows 11 devices.

As an upgraded version of MAI-Code-1-Flash released in June, MAI Code 1.1 Flash has now been integrated into development environments such as GitHub Copilot and VS Code for use in daily coding, repository Q&A, refactoring, tool invocation, and other scenarios.

Microsoft plans to further integrate it into the Windows system. This model uses 3-bit quantization, which has already reduced the size by 80%, and the context window size is 256K. Below are the relevant images from the IT community:

In terms of positioning, this model mainly helps you complete tasks efficiently, with high frequency, low cost, and low latency within Copilot. It plans, infers, and executes tasks in the actual development process, while striving to use as few tokens as possible and at a lower cost.

In terms of performance, compared to the June version, MAI-Code 1.1 - Flash has seen a improvement of about 22% on the basis of GitHub Copilot CLI's Terminal-Bench 2.1, and about 15% improvement on the NET task; at the same time, the token output speed has increased by about 25%, and the token required to complete the task has decreased by about 25%.

In terms of multimodality, a new capability has been added: "Write code based on images / Understand diagrams." This allows for direct input of images such as screenshots, architecture diagrams, and UI sketches, accelerating the process from design to implementation (the June version only supported plain text).