Copilot Fusion can now route between on-device and cloud models, and GitHub says the routing decision still happens in the cloud
AspireAI community · GitHub Changelog
GitHub shipped intelligent routing to local models into experimental preview, so Copilot Fusion can choose between supported on-device models and cloud models per request in Copilot CLI, the Copilot app and VS Code, without the developer switching models by hand. On supported Windows devices the MAI Code 1.1 Flash model is downloaded and configured through the Windows ML provider, with setup in the CLI's /model flow, a Model providers setting in the app, and a chat.copilot.localModel.enabled flag in VS Code. Bring-your-own is broader: Copilot CLI's local Fusion routing also works with manually configured models on Windows, macOS and Linux, serving MAI Code 1.1 Flash or Qwen3.8 Flash Next through a local OpenAI-compatible server, though the Copilot app's intelligent routing currently supports only managed Windows ML models.
The caveat GitHub states plainly is the interesting part: local inference does not make the session offline, because Copilot Fusion still makes its routing decisions in the cloud, and local models may carry different safety protections than the cloud models GitHub provides.
It is an opt-in experimental preview rather than a general release. The managed download path is Windows-only, bringing your own model means you own the download and the inference runtime and your device has to have the resources to run it, and adding a model through bring-your-own-key does not by itself make it eligible for Fusion routing.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.