Windows targets November 20 for first Aion Instruct rollout

Microsoft is changing the language model that runs directly on compatible Windows PCs. In a November 13 announcement, the company said Aion Instruct will replace Phi Silica, with the first Windows Insider Preview rollout targeted for November 20 and broader availability on retail devices targeted for late January 2027. The new model accepts up to 32,000 input tokens in a single request.

A Windows app that uses Phi Silica to summarize a passage or rewrite a sentence can keep offering that feature while Windows changes which model answers the request. Microsoft expects most apps to carry over without developer action. Apps that rely on custom instructions, their own model adaptations, or an exact output format are among the exceptions. A reply that reads normally on screen can still look different to the software that processes it.

Which PCs can run it

Aion Instruct has 2.3 billion parameters, down from 3.8 billion in Phi Silica. Its input limit rises from 3,500 to 32,000 tokens. A token is a piece of text used by the model, so these figures do not translate directly into a word count. An app can pass in longer material under the model's limit. The amount a user can submit also depends on what the app accepts and how it packages the request.

Microsoft plans to support Aion Instruct on compatible discrete graphics cards as well as on Copilot+ PCs with a neural processing unit, a chip designed for tasks such as running language models. The announcement names NVIDIA's RTX series and AMD's RDNA4 generation as examples and says hardware support will widen in stages. Phi Silica already runs on some graphics cards. Microsoft's current documentation for that model lists an experimental Windows build and compatible drivers among the conditions for GPU use. Those conditions describe the outgoing model.

A separate developer package is already available for eligible Snapdragon, Intel Lunar Lake and AMD Ryzen AI Copilot+ PCs. The sample app can show how long it takes for a reply to begin appearing. Microsoft says those measurements are preliminary and should not be read as the performance of the retail Windows release. The developer package runs on a supported neural processing unit and has no CPU fallback.

Exact text can change under the same interface

Apps built around a custom LoRA need a new adaptation for Aion Instruct. A LoRA is a small file trained to change how a base model responds for a particular task. Microsoft's sample documentation says the files trained for Phi Silica are incompatible with its replacement. Apps that look for specific phrases or formatting in a reply also have a reason to compare outputs before the switch.

Microsoft's Windows AI documentation describes another difference a user may never see. Several text features insert an invisible watermark into generated text. A reply can display the same characters to the reader while containing extra Unicode characters in the saved string. A program reading the reply as JSON or Markdown can interpret it differently or fail to read it, depending on where the marker appears. The documentation says a response may contain no marker, one, or several.

The two models overlap during the preview

Microsoft's developer documentation says Phi Silica will remain on Insider devices during the November rollout, with Windows controlling which model is active. During that period, developers can compare the two models on the same machine. The separate package available now gives developers an earlier look at the model under a more limited set of hardware and distribution conditions.

Microsoft aims to finish the transition with the January rollout to retail devices, at which point its documentation says Phi Silica will be removed. The company also targets January for removing the unlock token developers used to request access to the older model's features. Support for additional graphics configurations will expand as Microsoft validates them.