LanguageModel API can now run on devices with a compatible GPU. The rest of the release adds diagnostics for that path, updates the inference runtime, and fixes bugs in WinUI, windowing and ARM64EC startup.
GPU support for LanguageModel
The release notes say the language model can now run on devices with a compatible GPU. They don't name GPU makes, minimum video memory, or a Windows build requirement. This is not a promise that any PC with a graphics card will work.
Earlier experimental builds give some context. Microsoft's release notes for the experimental 2.2.2 build said the Language Model APIs run on non-Copilot+ PCs with a supported GPU, and listed NVIDIA GeForce RTX 30 series and newer with 6+ GB of vRAM as supported hardware. That experimental flight required Developer Mode and a Windows Insider Experimental Channel build. The 2.7.0 notes don't say whether those requirements and hardware limits still apply to the stable release. Check the current Phi Silica hardware documentation before committing to a support matrix.
Microsoft's Phi Silica guidance also sets expectations for GPU execution:
- On non-Copilot+ GPU machines, latency is higher and depends on GPU generation, VRAM and current GPU load.
- Prompt compression and speculative decoding are listed as available on Copilot+ NPU hardware but not on GPU.
- The model is downloaded in the background through Windows Update.
- Users manage installed models under Settings > System > AI Components, where the GPU model appears as "AI LanguageModel".
That guidance comes from the Phi Silica page, not the 2.7.0 notes. It is the closest official description of how the GPU path behaves.
Better diagnostics for developers
The release adds three kinds of information for apps.
- Adapter compatibility:
LanguageModelLowRankAdapterResult.Statusreports whether a LoRA adapter loaded cleanly or with warnings. It helps apps pick the right adapter. - Model identity:
LanguageModel.ModelNameandModelVersionlet an app see which model is installed. Microsoft says both identity and detailed compatibility results depend on the installed AI components. Code must tolerate those values being absent on some systems. - GPU resource errors:
LanguageModelResponseStatus.InsufficientVideoRamandBlockedByGameModetell an app when GPU inference stopped because video memory was low or Game Mode was active.
These are status values, not fixes. An app can use them to show a clear message. For example, it could tell the user to close a game, or fall back to another path. The platform does not clear the underlying condition for you. Handling the two new values separately from a generic failure is the sensible approach.
The experimental 2.5.4 notes show these ideas were in testing earlier. That build introduced LanguageModel.ModelName, ModelVersion and LanguageModelLowRankAdapterStatus. That experimental note names the low-VRAM status LowOnVram. The stable 2.7.0 notes call it InsufficientVideoRam. If you wrote code against the experimental builds, expect to rename it.
ONNX Runtime and Windows ML
ONNX Runtime moves to 1.28.3, and the Windows ML dependency moves to 2.5.96. For comparison, Microsoft's what's-new page says the October 27 release, 2.6.4, updated ONNX Runtime to 1.27.1. The notes give no performance or model-coverage figures for the newer versions. If you ship ONNX models, retest them rather than assuming the upgrade is harmless.
Bug fixes and their runtime-compatibility switches
Microsoft pairs most fixes with a runtime compatibility change name. Apps can use these to opt out of a fix if it causes problems.
| Fix | Compatibility change |
|---|---|
Input offset in DPI-aware apps when CompositionEngineType.System is selected | InputPointerSource_SystemEngineDPIOffset |
Text Cursor Indicator not showing immediately when an empty TextBox or RichEditBox gets focus | WinUI_TextCursorIndicatorForEmptyTextControl |
ContentDialog entrance animations not playing, and MenuFlyout items wrongly showing tooltips, when XamlChangeId.OptimizeApplyStyles is enabled | Xaml_OptimizeApplyStylesApplyTemplate |
Crash reading Window.DispatcherQueue from a background thread after the window's XAML thread shuts down | Window_DispatcherQueueAfterXamlShutdown |
LanguageModel streaming crash when the session host terminates, plus better recovery for eligible requests that fail before any response | LanguageModel_StreamingAndHostExitRecovery |
ARM64EC apps failing to start on ARM64 Windows because the runtime could not resolve the TransformPoint entry point | N/A (DLL import resolution happens before runtime containment can run) |
First-chance exceptions in default ScrollBar templates when DefaultStyleOptimizations is enabled | N/A (default template change in a .xaml file) |
A few of these have narrow scope.
- Window lifetime: The crash only occurs when a background thread reads
Window.DispatcherQueueafter the window's XAML thread has shut down. Apps with worker threads that outlive windows are the ones affected. - ARM64EC: This is a startup fix. The failing step is DLL import resolution, which happens before any compatibility opt-out could apply. That is why the table shows "N/A". If you ship ARM64EC builds, this is the item to retest first. It is tied to GitHub issue #6911.
- ScrollBar templates: The ScrollBar fix is a template change in a
.xamlfile, so it can't be switched off at runtime. It was a first-chance exception, which usually shows up in the debugger rather than as a crash. - Language model recovery: "Eligible requests" isn't defined in the notes. Keep your own retry and error handling.
- ContentDialog and MenuFlyout: The experimental 2.6.x line carried a related fix for the same GitHub issue (#11257). The same
Xaml_OptimizeApplyStylesApplyTemplatechange name appears in both, so the fix has moved from the experimental channel to stable.
How to try it
Install the Microsoft.WindowsAppSDK 2.7.0 NuGet package in your project, using Visual Studio's NuGet UI or your usual package workflow. Runtime deployment requirements still vary by app type, so adding the package does not by itself deploy every prerequisite for every packaged or unpackaged app.
A reasonable upgrade pass:
- Build and run your existing test suite on 2.7.0.
- If you use the language model, handle the two new response statuses explicitly.
- If you target ARM64EC, test startup on a real ARM64 Windows device.
- If something regresses, check whether the matching compatibility change name can be disabled while you investigate.
Context
Windows App SDK 2.0 shipped on April 29, 2026. It was the first major version since 1.0 and adopted Semantic Versioning. Under that scheme, the Language Model APIs no longer require apps to unlock a Limited Access Feature as of 2.6.4. That requirement removal, GPU execution and the new diagnostics together suggest Microsoft is making local language models easier to adopt beyond Copilot+ hardware.
2.7.0 is a developer-facing update. It changes nothing directly for end users. They will only see it through apps that adopt the new APIs and fixes. The practical points are that GPU inference depends on hardware and installed AI components, and that the new status codes should be handled explicitly.
References
- Windows App SDK 2.0 release notes - Windows apps | Microsoft Learn learn.microsoft.com
- Windows App SDK 2.0 release notes - Windows apps learn.microsoft.com