GitHub Copilot Fusion now supports selection of local models

GitHub Copilot Fusion can choose between supported local and cloud models in an opt-in experimental preview for Copilot CLI, the Copilot app and VS Code.

Managed setup uses MAI Code 1.1 Flash through Windows ML on supported Windows PCs with an NVIDIA GPU. It requires 96 GB of available combined RAM, 60 GB of storage and the documented llama.cpp setup. Once configured, select Standard in Fusion to use routing.

Copilot CLI also supports manually configured MAI Code 1.1 Flash or Qwen3.8 Flash Next on Windows, macOS and Linux through a local OpenAI-compatible server. Adding a model through BYOK alone doesn't make it eligible. The app's routing currently supports managed Windows ML models only.

Fusion still makes routing decisions in the cloud, so local inference doesn't make a Fusion session offline. GitHub also notes that local models may have different safety protections from its cloud models.