Microsoft sets out a local AI model to replace Phi Silica on Windows

What happened: Microsoft has outlined a move from Phi Silica to Aion Instruct, a language model with 2.3 billion parameters that runs on the device itself. It accepts up to 32,000 tokens of input. The company is extending it from Copilot+ PCs to supported GPUs, and a developer testing package is available now.
Why it matters: A local model can suit apps that need low delay or want to keep data on the device. Developers should test their prompts, custom LoRA adapters and output parsers again instead of treating the new model as a drop-in swap. Processing on the device alone does not mean an app works fully offline or sends no data.
Dates: The transition plan was announced on 13 November 2026. A rollout to Windows Insiders is planned for 20 November, and a wider retail rollout for January 2027, so the change is planned and not yet complete.