|
This article applies to Copilot+ PCs only. |
Introduction
Phi Silica is a Windows AI component for AMD‑powered Copilot+ PCs that enables on‑device language intelligence across Windows features and applications. It is a Microsoft‑developed, Transformer‑based small language model (SLM) optimized for machine‑learning inference on the device’s Neural Processing Unit (NPU), allowing language tasks to be executed entirely on the device rather than in the cloud. This on‑device execution delivers fast, low‑latency responses while keeping user data local to support privacy‑focused workflows. Despite its compact footprint, Phi Silica provides core language model capabilities such as text understanding, summarization, rewriting, and short‑form content generation within the power and performance constraints of Copilot+ hardware. The Phi Silica AI component supports a broad range of Windows and Copilot+ experiences—including productivity, accessibility, and AI‑assisted workflows—and is exposed to developers through Windows AI APIs, enabling applications to integrate local, machine‑learning–based language processing without requiring cloud connectivity.
Summary
This Phi Silica update is a new release for the Phi Silica AI component for Windows 11, version 24H2 and Windows 11, version 25H2.
How to get this update
Windows Update
This update will be downloaded and installed automatically from Windows Update.
Prerequisites
You must have the latest cumulative update for Windows 11, version 24H2 or Windows 11, version 25H2 installed.
Replacement information
This update replaces the previously released update KB5084167.
Check update history
To determine whether this update is present on the device, go to Settings > Windows Update > Update history.
After this update is installed, you should find the following update listed in Update history depending on the processor type in the device:
-
2026-04 Phi Silica version 1.2604.515.0 for AMD-powered systems (KB5090933)
References
For a list of AI updates on supported devices, see the following resources:
