AI software on a Windows PC runs a model file on the processor, the graphics chip, or a neural processing unit (NPU), and Microsoft's current runtime for that job is Windows ML, which is built on ONNX Runtime. DirectML, the older graphics-based layer, is in what Microsoft calls sustained engineering: still supported, but no longer receiving new features. Which program suits a given person depends on the task and the hardware, so this page covers the parts that decide whether an AI feature will work on a particular PC instead of ranking products.
Key takeaways
- Windows ML runs ONNX models on CPU, GPU, or NPU through execution providers, and the hardware-specific providers need Windows 11 version 24H2 (build 26100) or later.
- Microsoft's documentation says DirectML is in sustained engineering and points new work to Windows ML.
- A Copilot+ PC is not required for Windows ML or Foundry Local, but most of the built-in Windows AI APIs expect one.
- A feature marketed as local can still fall back to a cloud service, a pattern Microsoft itself documents.
- Where an app gets its model file, and whether it checks the download, are fair questions for any vendor.
What "runs locally" means on a Windows PC
An AI program is a normal application plus a model, which is a large file of numbers learned during training. At run time, a runtime library loads the file and carries out the math. On Windows, that library is usually ONNX Runtime, an open standard for model files, and Microsoft ships its own supported copy as Windows ML. Models trained in PyTorch, TensorFlow, or scikit-learn can be converted to the ONNX format first.
The runtime hands the math to one of three kinds of silicon. The CPU is the universal fallback. The GPU handles high-throughput work such as images and video. The NPU is a dedicated chip meant for sustained AI work at low power, and Microsoft's documentation says the most powerful NPUs are found in Copilot+ PCs. The practical consequence is that the same app can feel quick on one laptop and sluggish on another without any change in the software.
Four Microsoft layers and where each stands
Microsoft has renamed its Windows AI platform several times. Its own terminology table says "Copilot Runtime APIs" was a 2024 name, "Windows AI Foundry" a 2025 name, and "Microsoft Foundry on Windows" is the current umbrella for three parts: the Windows AI APIs, Foundry Local, and Windows ML. The cloud service called Microsoft Foundry is a separate product. The table below follows the status given on Microsoft Learn as of October 2026.
| Layer | What it does | Hardware | Status in Microsoft's docs |
|---|---|---|---|
| Windows AI APIs | Ready-made models for text, images, and similar tasks, including Phi Silica | Most need a Copilot+ PC; a few also run on other Windows 11 hardware | Current |
| Foundry Local | Runs open-source language and speech models through an OpenAI-compatible API | Any Windows hardware, Windows 11 24H2 or later for the SDK | Current; the model catalog changes over time |
| Windows ML (new) | Runs any ONNX model with hardware acceleration | x64 or ARM64, any configuration | Current, actively developed |
| Windows ML (old) and DirectML | Earlier inference API (inbox since Windows 10 version 1809) and a low-level DirectX 12 layer | DirectX 12 GPUs for DirectML | Old API: no new investment. DirectML: sustained engineering |
The comparison page also states that a Copilot+ PC is defined by hardware, namely an NPU rated at 40 or more TOPS and at least 16 GB of memory. It is not needed for Foundry Local or Windows ML.
What the hardware-specific providers require
Windows ML loads optional execution providers, one per chip family, and Windows delivers them through Windows Update. The documentation lists AMD (MIGraphX and VitisAI), NVIDIA (TensorRT for RTX), Intel (OpenVINO), and Qualcomm (QNN), plus an experimental WebGPU provider. The fine print matters to buyers and users:
- NVIDIA's provider lists GeForce RTX 30 series and newer.
- AMD's MIGraphX provider lists RDNA 3 or later GPUs with a stated minimum driver, and says it is not supported for generative AI scenarios.
- Qualcomm's provider lists Snapdragon X Elite and X Plus with a Hexagon NPU driver of version 30.0.140.0 or later.
- Intel's driver requirements are on a separate Intel page that Microsoft links.
The CPU and DirectML providers come built in. Everything above, per the documentation, applies to Windows 11 24H2 or later. A PC on an older Windows 11 release can still run models on the CPU and, through DirectML, the GPU, which explains why an AI feature sometimes works on an older machine but without NPU speed-ups.
Where the model file comes from
Windows ML loads a model from a local file path, and Microsoft describes two ways to put the file there. An app can bundle the model in its package, which suits small models and guarantees offline use from the first launch. Or it can download the model later, which keeps the installer small for models that run to gigabytes. For downloads, the documentation tells developers to host the file over HTTPS, publish a SHA-256 hash, verify it before use, and write the file to a temporary name first so an interrupted download never leaves a corrupt model behind. Foundry Local's quickstart notes that the first run needs internet access for the package, model, and runtime downloads.
A local label can still involve the cloud
Microsoft's guide to choosing a Windows AI option describes a three-tier pattern for developers: try the Windows AI APIs first, fall back to Foundry Local on any hardware, and fall back to Azure AI in the cloud if neither is available. Nothing is wrong with that design, and it keeps a feature working on every PC. It does mean that "uses the NPU" and "never sends data off the device" are different claims. The second one has to come from the vendor's privacy documentation, not from the presence of local inference.
Questions to put to an AI program before installing it
- Which tasks run on the device, and which are sent to a server?
- Which hardware and Windows version does the local mode need, and what happens on a PC that lacks it?
- Is the model bundled or downloaded, how large is the download, and does the app ask first?
- Does the vendor say how it verifies downloaded model files?
- Does the app keep working offline once the model is on the disk?
- If an execution provider from a chip maker is used, are its license terms listed? Microsoft publishes third-party notices for each provider.
The old page title promised "top" software. A ranking would need controlled tests on specific tasks and hardware, and none are cited here, so no product is named as best. This article gives general information and is not a security audit or a purchase recommendation.
FAQ
Do AI programs on Windows need a Copilot+ PC?
No. Microsoft's documentation says Windows ML and Foundry Local run on any Windows hardware, with speed depending on the CPU, GPU, or NPU present. Most of the built-in Windows AI APIs do expect a Copilot+ PC.
Is DirectML discontinued?
No. Microsoft's page says it continues to be supported, but new feature development has moved to Windows ML. It was introduced with Windows 10 version 1903.
What is an NPU used for?
It is a chip meant for sustained AI inference at low power draw. Microsoft's description says NPUs are battery-efficient, while GPUs generally give the highest throughput on discrete cards.
Can a local AI feature work with no internet connection?
Once the model file is on the disk, a bundled or already downloaded model can run offline, according to Microsoft's model distribution guide. The first download needs a connection, and an app that falls back to a cloud service needs one for that path.
Related reading
- Copilot+ PC requirements explained
- what the AI Host process is
- software, hardware, and model files
- how testing differs for AI programs
More in AI Tools and Industry.
Last verified: October 2026
Sources
- What is Windows ML?, Microsoft Learn, updated September 2026
- Windows ML execution providers, Microsoft Learn, updated October 2026
- Distribute models for Windows ML, Microsoft Learn, 2026
- Introduction to DirectML, Microsoft Learn, 2026
- Choose your Windows AI solution, Microsoft Learn, 2026
- Get started with Foundry Local, Microsoft Learn, 2026