Microsoft · 7 October 2026

Microsoft put a smaller coding model in Copilot, and on the machine.

Summary of Microsoft’s own post. The claims below are theirs.

Microsoft says MAI-Code-1.1-Flash is in production in GitHub Copilot. Against the coding model they launched in June, they say this one writes better code, uses 25 percent fewer tokens, and costs a quarter as much. Inside Copilot they say tokens stream 25 percent faster. They also cite a 22 percent gain on a command-line test in Copilot CLI, and a 15 percent gain on .NET tasks. Code that people kept rose 4 percent, and return visits rose 9 percent, by their count.

They also say a 3-bit version keeps a 256,000-token window and stays close to the full-size model on the coding tests they name. Local calls, they say, are not billed. They recommend a machine with more than 120GB of memory. Experimental access in the Copilot app, the Copilot CLI, and Visual Studio Code was promised by the end of October.

The builder is moving onto the computer you already own.

BitBytec assumes that. The agent is one you already run. The difference is the front door: a blueprint, a yes on your Mac, and a preview. We do not claim this Microsoft model is the one inside the IDE.