Hacker News · AI· doomroot13·· 7 hours agoSelectedAI score71
Microsoft Unveils MAI-Code-1.1-Flash: A 137.9B Code Model Brought to On-Device Inference on Windows PCs
Advancing the Quality-Efficiency Frontier for Coding AI on Windows PCs
AI guide
Microsoft has released MAI-Code-1.1-Flash, a 137.9B MoE code model, compressing its weights to 53.0 GiB via mixed-precision quantization to serve as a local, on-device model for GitHub Copilot.
Why it matters
The article highlights the co-design of mixed-precision quantization, speculative decoding, and long-context KV cache management, offering an engineering blueprint for the on-device deployment of large-parameter MoE models.
Source:Hacker News · AI · devblogs.microsoft.com