Which Local AI Features Are Now Open

The expanded APIs are part of the Windows Copilot Runtime and give applications direct access to on-device AI processing. In this public preview, Windows 11 PCs with a supported RTX GPU can run the following workloads locally :

• Text Summarization and Rewriting: Apps can call a local language model to condense or rephrase content entirely offline.
• Text Generation: Developers can integrate local small language model (SLM) capabilities into their Windows applications.
• Text-to-Image Generation: The API path supports local image creation without a cloud roundtrip .
• Video Super-Resolution and Speech Recognition: These specific features run on the CPU in the current preview, but they are part of the same expanded API set that no longer requires an NPU .

It is critical to note that this is a developer-facing release delivered via the Windows App SDK, not a consumer-facing toggle or a standard Windows Update. The feature is explicitly marked "Experimental" in Microsoft's documentation .

A Full Strategic Pivot Away from NPU Exclusivity

Since the launch of Copilot+ PCs, Microsoft maintained a strict rule: the native Windows Local Language Model APIs and the Copilot+ branding features were available only on systems with an integrated NPU . The NPU was the only recognized AI accelerator for Microsoft's local intelligence stack.

This latest release officially breaks that seal. An Nvidia RTX GPU can now serve as the primary AI accelerator for Windows' local language models, a role previously reserved exclusively for NPU silicon. The industry's verdict was swift and direct. PCWorld described the move as one that "chips away at the Copilot+ advantage," while PCMag went further, stating that the message from Microsoft's Build 2026 conference was that "Copilot+ PCs no longer matter" for local AI development .

The company is no longer gating its native AI runtime behind a specific hardware badge. Any Windows 11 system with a capable RTX GPU can now participate in the local AI ecosystem, a change that dramatically widens the target market for AI-powered Windows applications.

The Fine Print: Current Limitations of the GPU Preview

Despite the strategic significance, this GPU-based support remains a work in progress with several important limitations:

• Developer Preview Only: This is not a stable, generally available release. It is intended for developers experimenting with local model integration via the App SDK, not for end users seeking a new feature drop .
• No Copilot+ Exclusive Features: Accessing the local Language Model APIs via an RTX GPU does not unlock Copilot+ branded features like Windows Recall, Click to Do, or Windows Studio Effects. Those experiences retain their NPU dependency .
• Phi Silica Not Yet Supported: Reports indicate that Microsoft's flagship on-device SLM, Phi Silica, is not yet optimized for the RTX GPU path; it remains configured for NPU execution in this preview phase .
• Strict Hardware Floor: The minimum requirement is an RTX 30-series or newer GPU with at least 6GB of VRAM. Older GTX cards, RTX 20-series GPUs, and any card with less VRAM are excluded .
• Developer-Only Focus: The APIs are strictly plumbing for app integration, not a user-accessible interface or feature toggle .

The New Hardware Anchor: Surface RTX Spark Dev Box

Microsoft paired this software pivot with a powerful hardware statement at Build 2026. The company unveiled the Surface RTX Spark Dev Box, a desktop-class developer workstation built on Nvidia's new RTX Spark system-on-a-chip, which combines a Grace-class Arm CPU with a Blackwell RTX GPU on a single package .

Key specifications for the RTX Spark platform include :

• 1 petaflop of AI compute performance
• Up to 6,144 Blackwell RTX cores and 20 power-efficient Arm cores
• 128GB of unified system memory, capable of running large language models with up to approximately 120 billion parameters entirely on-device

Microsoft promotes the Dev Box as a "desktop data center" for local AI development, a form factor designed to make large-model inference and agentic workflows practical without cloud dependency . The device follows the earlier announcement of the Surface Laptop Ultra, also RTX Spark-powered, creating a family of Windows on Arm devices built from the ground up for local AI . The broader developer toolchain includes native TensorRT for RTX integration within Windows ML and a DGX Station for Windows ecosystem .

By opening the local AI runtime to discrete GPUs and simultaneously launching Nvidia-powered first-party hardware, Microsoft is re-centering its developer strategy on GPU performance rather than on a specific processor type. For anyone building on-device AI features for Windows, the path forward is now defined by raw compute and memory capacity, not by an NPU sticker.

Which Local AI Features Are Now Open

The expanded APIs are part of the Windows Copilot Runtime and give applications direct access to on-device AI processing. In this public preview, Windows 11 PCs with a supported RTX GPU can run the following workloads locally :

• Text Summarization and Rewriting: Apps can call a local language model to condense or rephrase content entirely offline.
• Text Generation: Developers can integrate local small language model (SLM) capabilities into their Windows applications.
• Text-to-Image Generation: The API path supports local image creation without a cloud roundtrip .
• Video Super-Resolution and Speech Recognition: These specific features run on the CPU in the current preview, but they are part of the same expanded API set that no longer requires an NPU .

It is critical to note that this is a developer-facing release delivered via the Windows App SDK, not a consumer-facing toggle or a standard Windows Update. The feature is explicitly marked "Experimental" in Microsoft's documentation .

A Full Strategic Pivot Away from NPU Exclusivity

Since the launch of Copilot+ PCs, Microsoft maintained a strict rule: the native Windows Local Language Model APIs and the Copilot+ branding features were available only on systems with an integrated NPU . The NPU was the only recognized AI accelerator for Microsoft's local intelligence stack.

This latest release officially breaks that seal. An Nvidia RTX GPU can now serve as the primary AI accelerator for Windows' local language models, a role previously reserved exclusively for NPU silicon. The industry's verdict was swift and direct. PCWorld described the move as one that "chips away at the Copilot+ advantage," while PCMag went further, stating that the message from Microsoft's Build 2026 conference was that "Copilot+ PCs no longer matter" for local AI development .

The company is no longer gating its native AI runtime behind a specific hardware badge. Any Windows 11 system with a capable RTX GPU can now participate in the local AI ecosystem, a change that dramatically widens the target market for AI-powered Windows applications.

The Fine Print: Current Limitations of the GPU Preview

Despite the strategic significance, this GPU-based support remains a work in progress with several important limitations:

• Developer Preview Only: This is not a stable, generally available release. It is intended for developers experimenting with local model integration via the App SDK, not for end users seeking a new feature drop .
• No Copilot+ Exclusive Features: Accessing the local Language Model APIs via an RTX GPU does not unlock Copilot+ branded features like Windows Recall, Click to Do, or Windows Studio Effects. Those experiences retain their NPU dependency .
• Phi Silica Not Yet Supported: Reports indicate that Microsoft's flagship on-device SLM, Phi Silica, is not yet optimized for the RTX GPU path; it remains configured for NPU execution in this preview phase .
• Strict Hardware Floor: The minimum requirement is an RTX 30-series or newer GPU with at least 6GB of VRAM. Older GTX cards, RTX 20-series GPUs, and any card with less VRAM are excluded .
• Developer-Only Focus: The APIs are strictly plumbing for app integration, not a user-accessible interface or feature toggle .

The New Hardware Anchor: Surface RTX Spark Dev Box

Microsoft paired this software pivot with a powerful hardware statement at Build 2026. The company unveiled the Surface RTX Spark Dev Box, a desktop-class developer workstation built on Nvidia's new RTX Spark system-on-a-chip, which combines a Grace-class Arm CPU with a Blackwell RTX GPU on a single package .

Key specifications for the RTX Spark platform include :

• 1 petaflop of AI compute performance
• Up to 6,144 Blackwell RTX cores and 20 power-efficient Arm cores
• 128GB of unified system memory, capable of running large language models with up to approximately 120 billion parameters entirely on-device

Microsoft promotes the Dev Box as a "desktop data center" for local AI development, a form factor designed to make large-model inference and agentic workflows practical without cloud dependency . The device follows the earlier announcement of the Surface Laptop Ultra, also RTX Spark-powered, creating a family of Windows on Arm devices built from the ground up for local AI . The broader developer toolchain includes native TensorRT for RTX integration within Windows ML and a DGX Station for Windows ecosystem .

By opening the local AI runtime to discrete GPUs and simultaneously launching Nvidia-powered first-party hardware, Microsoft is re-centering its developer strategy on GPU performance rather than on a specific processor type. For anyone building on-device AI features for Windows, the path forward is now defined by raw compute and memory capacity, not by an NPU sticker.