What Thurrott reports
Thurrott's article, "Windows 11, RTX Spark, and Local AI: It's Not Me, It's Windows," argues that local AI models are ordinary files that sit in the file system, yet they don't behave like other shared files on Windows. The article is behind a Premium paywall, and the public portion cuts off mid-sentence while he discusses cleanup. I can only report what appears in that public excerpt.
His main points:
- Windows normally offers a shared file system, hardware abstraction through drivers, and files that any app can open. A downloaded DOCX isn't locked to Word.
- He says local AI models don't follow that pattern. There's no formal storage location and no system service acting as a middleman between models and the software that needs them.
- In practice, you can install an AI app such as Hermes, LM Studio, Clairvoyance or Windows Llama, and download a model through its built-in browser. A second app can't see that model, so using the same model there means downloading it again.
- RTX Spark machines make this worse. They can have up to 128 GB of memory, and model sizes keep growing. Experimenting with different apps, runtimes and models fills the disk with leftovers.
- He also says Microsoft claims over 40 local AI models ship on every Copilot+ PC. They're hidden from the user and updated through Windows Update or the Microsoft Store, depending on the model.
This is Thurrott's own experience. The excerpt doesn't give specific settings, model formats or file paths. It also doesn't show whether any of those apps can be pointed at a shared model folder. Nor can I confirm the "over 40 models" figure from Microsoft's documentation.
Windows does have a model-sharing API
Thurrott says there is no system service for sharing models. Microsoft's developer documentation describes something close to one: the Windows ML Model Catalog APIs.
These APIs download models from an online catalog and store them locally so that any Windows application on the device can use them. Microsoft lists the core features as adding catalogs, finding models compatible with the user's hardware, downloading them, and sharing them across apps. The sharing rule is specific: if several applications request the same model with the same SHA-256 hash, they share the copy already on the device. Downloaded models go to a user-specific location.
The setup for developers:
- They need Windows App SDK 1.8.3 or later and at least one model catalog source.
- A catalog source is a JSON file describing available models, their compatibility and download details. It can be hosted at an https:// address or referenced as a local file.
- The app searches the catalog by model name, and the catalog picks a variant that matches the device's hardware.
- When the app requests the model, it's downloaded if needed, and the app receives a file path to use in its own inference code.
Microsoft also says who the feature is for. The documentation describes it as aimed at library and SDK authors who want to download a model from a CDN and share it across every app that uses their library. If several independent apps depend on a library and should share one on-disk copy of a model, Microsoft says to use the Model Catalog APIs rather than writing custom download-and-share logic.
So the mechanism exists. The catch is adoption. Nothing in the documentation I reviewed shows that Hermes, LM Studio or the other apps Thurrott names use it. If they manage their own downloads, they won't benefit. In that sense his experience and Microsoft's documentation can both be true.
Three kinds of local AI model on Windows
Part of the confusion is that "local AI models on Windows" covers several different things.
| Category | How models arrive | Sharing and cleanup |
|---|---|---|
| Windows AI APIs (Phi Silica, OCR, image features) | Microsoft handles distribution and runtime | Microsoft describes them as hosted by Microsoft, acquired at runtime, and shared across apps. |
| Foundry Local | Microsoft describes it the same way: hosted by Microsoft, acquired at runtime, and shared across apps. | Aimed at developers and open-source LLMs |
| Windows ML (bring your own model) | Distribution is handled by your app, though app libraries can share models. | Depends entirely on the app or library |
Third-party model browsers such as those in LM Studio sit outside this table. Their sharing and cleanup behavior depends on each app.
Where Windows does give you a cleanup path
For some Microsoft-managed components, Windows offers a place to manage them. Phi Silica is preinstalled on Copilot+ PCs. On GPU and CPU devices it's downloaded on demand through Windows Update and can run to several gigabytes. Microsoft's Windows AI APIs documentation says users can remove or reinstall it at Settings > System > AI Components. Image Generation and, on CPU-only devices, Speech Recognition are handled the same way. Video Super Resolution is different: it ships with the Windows App SDK and isn't a removable model.
This matches part of Thurrott's complaint. Microsoft's own models have a defined lifecycle, but none of the sources I reviewed describe a Settings page that lists and removes every model downloaded by every third-party AI app.
Foundry Local is a separate path
Foundry Local is Microsoft's runtime for running LLMs locally on Windows. Its quickstart lists these requirements:
- Windows 11 24H2 (build 26100) or later
- .NET 9 SDK or later
- An x64 or Arm64 device with enough memory and disk for the chosen model
- Internet access for the initial package, model and runtime downloads
A dedicated GPU, NPU or Copilot+ PC isn't required if the model has a compatible CPU variant. The package for Windows is Microsoft.AI.Foundry.Local.WinML. When you request a model by alias, Foundry Local can pick a QNN NPU variant on Snapdragon, a CUDA variant on NVIDIA, or a CPU variant. The optional command-line tool installs with winget install Microsoft.FoundryLocal, and foundry model list shows the available aliases.
Microsoft's quickstart also warns that Foundry Local aliases and recommended models change as the catalog evolves. Don't treat its phi-4-mini example as permanent.
Nothing here shows that Foundry Local shares its downloads with the third-party apps Thurrott mentions.
Why RTX Spark makes the problem more visible
NVIDIA announced on October 16, 2026 that RTX Spark laptops were available from ASUS, Dell, HP, Lenovo, Microsoft and MSI. Its announcement describes a chip pairing a Blackwell RTX GPU with a Grace CPU. It cites up to 128 GB of unified memory and up to 1 petaflop of FP4 AI performance. NVIDIA says the systems are meant to run large language models and AI agents locally, and it lists Hermes and OpenClaw among day-one agents. Its stated starting price is $2,599, and it says pricing and configurations vary by region.
Those are NVIDIA's own claims, not independent benchmarks. There is also a distinction worth keeping straight:
- Unified memory is the pool the model runs in. It determines how large a model you can run.
- Disk storage is where duplicate model files pile up.
A big memory pool encourages people to try bigger models, and bigger models mean bigger files. Thurrott's complaint is about the second problem, which the first one makes worse.
What to do about duplicate models
I can't verify app-specific settings from these sources, so this is general guidance rather than a tested procedure.
- Check before you download again. Look in each app's settings for a model directory or a way to import existing model files. Some apps may offer this, but I haven't confirmed which.
- Know your formats. Differences in model format, quantization, runtime and metadata can stop one app from using another's files. This is general technical background, not a confirmed cause in Thurrott's case.
- Don't hand-delete unknown model files. Use the owning app's own uninstall or delete function where it exists, and check Settings > System > AI Components for Microsoft-managed models.
- Developers: consider the Model Catalog APIs if you ship a library that several apps depend on. Show users which models are installed, where they are and how to remove them.
The bottom line
Thurrott is right that the user experience is fragmented. If you juggle several local AI apps on a new Windows PC, you're likely to hit duplicate downloads and unclear cleanup. His claim that Windows has no sharing mechanism at all goes too far, though. Microsoft documents a hash-based sharing system in the Model Catalog APIs. It's aimed at developers and only helps if app makers adopt it.
Whether that changes depends on whether popular local AI apps start using the Windows catalog or stay with their own downloaders. Until then, the practical advice for RTX Spark owners is to treat model storage like any other resource: check what you already have, know where it lives, and delete only through the app that put it there.
References
- Windows 11, RTX Spark, and Local AI: It’s Not Me, It’s Windows ⭐ - Thurrott.com Thurrott.com · 2026-10-23T21:44:12+00:00
- NVIDIA RTX Spark PCs Available Today — A New Beginning for Windows PCs | NVIDIA Newsroom nvidianews.nvidia.com
- Get started with Windows ML Model Catalog APIs learn.microsoft.com