MAI-Code-1.1-Flash gets a local build for Copilot, with a 120GB+ RAM recommendation

Microsoft says MAI-Code-1.1-Flash now runs locally with no per-call charges, and Copilot's router will be able to use it from late October.

Microsoft has released an on-device build of MAI-Code-1.1-Flash, the small coding model already in production in GitHub Copilot. Microsoft AI's October 7 post says local model calls carry zero inference charges. Once the integration ships, Copilot's router will be able to send eligible coding work to your machine and keep cloud models for tasks that need more capability. If you run agents all day, that changes the cost math. The catch is hardware: Microsoft recommends more than 120GB of RAM for best performance.

According to Microsoft, the hosted 1.1 writes better code than the 1.0 model launched at Build, uses 25% fewer tokens per task and costs a quarter as much. The local build uses 3-bit quantization and keeps the full 256K context. AICoder reports that Microsoft measured 70.80% on SWE-Bench Verified and 66.29% on Terminal-Bench 2.1 locally, with 75.5GB peak memory at 256K context. These are vendor numbers, and none of the sources found an independent reproduction. WindowsForum notes that 120GB is a best-performance recommendation, not a published minimum. It also found no compatibility list for machines other than RTX Spark PCs.

The weights are downloadable now, Microsoft says. The Copilot integration is not. Experimental access in the Copilot app, Copilot CLI and VS Code is due by the end of October. It's still unclear where the weights are hosted and what license covers them.

If your machine has less memory, there are routes available today. XenoSpectrum reports that Copilot CLI 1.0.94-0 and later detects Ollama models via /model, as long as they support tool calling and streaming. XenoSpectrum says pointing Copilot at a local OpenAI-compatible endpoint comes with the update whose preview starts in late October. Nobody has published a head-to-head benchmark against MAI-Code. Microsoft also showed DeepSeek V4 Flash running on RTX Spark at the same event.

Sources