Language
Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Windows AI APIs are the ready-to-use model APIs in Microsoft Foundry on Windows. They let you add supported features without finding, packaging, or optimizing your own machine learning model. These APIs run models locally on supported Windows 11 devices; the required Windows build, SDK release channel, and NPU, GPU, or CPU capabilities vary by API.
Supported hardware
Windows AI APIs are expanding beyond Copilot+ PCs to support a broader range of hardware. The following table shows the current hardware support for each API.
Note
On a Copilot+ PC, supported APIs generally use the NPU by default. Individual APIs can have different selection and fallback behavior. For Video Super Resolution, automatic selection doesn't guarantee NPU execution, and experimental creation options support explicit device requests. See the individual API pages for SDK availability and hardware requirements.
| API | NPU (Copilot+ PC) | GPU | CPU |
|---|---|---|---|
| Phi Silica | ✅ Available | ✅ Available (NVIDIA and AMD) | ❌ Not supported |
| Text Recognition (OCR) | ✅ Available | ❌ Not supported | ❌ Not supported |
| Speech Recognition | 🧪 Experimental | ❌ Not supported | 🧪 Experimental (optional, removable) |
| Video Super Resolution | ✅ Available | Experimental option (contract 2) | ✅ Available |
| Image Super Resolution | ✅ Available | ❌ Not supported | ❌ Not supported |
| Image Description | ✅ Available | ❌ Not supported | ❌ Not supported |
| Image Segmentation | ✅ Available | ❌ Not supported | ❌ Not supported |
| Object Erase | ✅ Available | ❌ Not supported | ❌ Not supported |
| Image Generation | ✅ Available (optional, removable) | ❌ Not supported | ❌ Not supported |
Note
GPU support for Phi Silica is available on NVIDIA GeForce RTX 30 series and newer (6+ GB vRAM) and AMD Radeon RX 9060 series and newer (6+ GB vRAM). GPU inference requires Developer Mode to be enabled (Settings > System > For developers) and the latest GPU driver installed directly from the manufacturer (see Phi Silica — GPU driver requirements). Video Super Resolution and Speech Recognition offer CPU execution and perform best on devices that meet the recommended specifications (4 physical cores, 3 GHz or higher base clock, 32 MB or more of L3 cache). See the individual API pages for details and a runtime check.
Model availability
The way the underlying AI model reaches a device depends on the API:
- Phi Silica — On Copilot+ PCs the model is preinstalled on the NPU. On GPU and CPU devices the model is not preinstalled — it is downloaded on demand the first time your app calls
EnsureReadyAsync. Downloads can be several GB and run in the background through Windows Update. End users can remove or reinstall the model at Settings > System > AI Components. Apps should checkGetReadyStatefirst and show a consent dialog before triggering the download. See Phi Silica — Model availability and download for the recommended UX pattern. - Image Generation — Runs on the NPU only, but the model is not preinstalled because of its install size. It is downloaded on demand the first time your app calls
EnsureReadyAsync, and users can later remove it at Settings > System > AI Components. Apps should checkGetReadyStatefirst and show a consent dialog before triggering the download. See Image Generation — Model availability and download for the recommended UX pattern. - Video Super Resolution — Apps use the SDK's deployment model for required runtime and model assets. Readiness calls are optional capability checks, not model downloads or device preparation, and don't guarantee successful creation or frame processing. See Create a VideoScaler session.
- Speech Recognition — On Copilot+ PCs the model is preinstalled on the NPU. On CPU-only devices the model is not preinstalled — it is downloaded on demand the first time your app calls
EnsureReadyAsync, and users can later remove it at Settings > System > AI Components. Apps should checkGetReadyStatefirst and show a consent dialog before triggering the download on CPU. See Speech Recognition — Model availability and download for the recommended UX pattern.
See the Windows AI APIs with WinUI sample app for how to use Microsoft Foundry on Windows with WinUI.
SDK and Windows availability
Don't use a historical feature-introduction version as a recommendation for the package you install today. Select a supported release from Windows App SDK release channels, then check the prerequisites for each API and overload you use. That page links to the current release notes and servicing policy.
| Capability | Feature-specific availability and prerequisites |
|---|---|
| Language models and text intelligence | Phi Silica, including model transition, hardware, driver, access, and Windows build requirements |
| LoRA adaptation | Phi Silica LoRA, including required SDK and model support |
| Image APIs | Imaging overview and each linked API page |
| Text recognition | Text recognition |
| Speech recognition | Speech recognition, including experimental API and CPU/NPU requirements |
| Video Super Resolution | Prerequisites and API availability, including overload-specific contract requirements |
| Semantic search | App Content Search, including access and SDK prerequisites |
Some features require a preview or experimental SDK and a Windows Insider build; see the release-channel differences. A newer stable package doesn't automatically include experimental APIs, grant restricted-feature access, or make an unsupported device compatible. Check runtime availability and model readiness before exposing a feature in your UI.
Note
Video Super Resolution's creation-options overload and input color properties belong to the experimental contract version 2 surface, not the baseline API availability listed above. Confirm the exposed members and Windows prerequisites for your selected SDK release. See Prerequisites and API availability.
Build your first app with Windows AI APIs
Tip
To improve accessibility and readability, this page displays still images by default. In some cases, you can click an image to see an animated version.
To build your first Windows app with these APIs, review the prerequisites and follow the example code in Get started building an app with Windows AI APIs.
You can then follow focused tutorials for the Phi Silica walkthrough, imaging walkthrough, and text recognition walkthrough.
Try the APIs and models on your PC
AI Dev Gallery is a demo app—available from the Microsoft Store—that lets you quickly download, try out, and use Windows AI APIs and models.
In AI Dev Gallery, select the Windows AI APIs tab menu item, then select the Phi Silica sample. If the model is already available on your device, then that sample will run straight away. Otherwise, select Request model to download the model. Once downloaded, that sample will be activated. Learn more about the AI Dev Gallery in What is the AI Dev Gallery?.
Overview of available APIs
Here are some ready-to-use capabilities that you can add to your Windows app:
Phi Silica
Important
Phi Silica is being replaced by Aion Instruct, a new on-device model. Aion Instruct begins rolling out to Windows Insider Preview devices in December 2026 and to retail devices in January 2027, at which point Phi Silica will be removed. See Get started with Phi Silica for transition details and timeline.
Similar to Large Language Models (LLM), Phi Silica is a Small Language Model (SLM) developed by Microsoft Research to perform language-processing tasks on a local device (see Get started with Phi Silica). Phi Silica is designed for Windows devices with a Neural Processing Unit (NPU) or a supported GPU, allowing text generation and conversation features to run in a high performance, hardware-accelerated way directly on the device. Phi Silica is not available in China.
Text recognition
The text recognition APIs recognize text in images and convert scanned documents, PDF files, and camera images into editable and searchable data on the device. See Get started with text recognition.
Imaging
Scale and sharpen images (Image Super Resolution), identify objects within an image (Image Object Extractor), generate natural-language descriptions of images (Image Description), and remove objects from images (Object Erase). See the imaging overview.
Image Super Resolution
The Image Super Resolution APIs enable image sharpening and scaling.
Also see Image Super Resolution.
Image Object Extractor
The Image Object Extractor APIs enable identifying objects within images.
Also see Image Object Extractor.
Image Description
The Image Description APIs describes images in natural language.
Note
Image Description features are not available in China.
Also see Image Description
Object Erase
You can use the Object Erase APIs to remove objects from images.
Also see Object Erase
Planned features
The following capabilities are planned but not yet available for use:
- Live Translation (Not yet supported). Help everyone using Windows—including those who are deaf or hard of hearing—better understand audio by viewing captions of spoken content (even when the audio content is in a language that's different from the system's preferred language).
Content moderation
Learn how content is moderated by the Windows AI APIs, and how you can adjust sensitivity filters. See Content safety moderation with the Windows AI APIs.
When adding model-powered features, review Responsible Generative AI Development on Windows.
Additional resources
- Code samples and tutorials. Explore ways to add local models and Windows APIs to your apps.
- Integrate AI in enterprise apps using Windows AI APIs. Watch the demo session from the November 2024 Microsoft Ignite conference.
- Provide feedback on these APIs and their functionality by creating a new Issue in the Windows App SDK GitHub repo or by responding to an existing issue.