AI Dev Gallery: a Windows sample browser for local AI models

The gap AI Dev Gallery fills for Windows developers

The hard part of adding a local model to a Windows app is rarely the model call itself. It is knowing which runtime to load, how to get the model weights onto the machine, how to bind the output to a UI, and what the whole thing looks like when it is assembled. AI Dev Gallery is Microsoft's answer to that first step. It is a WinUI 3 desktop application that ships over 25 interactive samples, each backed by a local model, with the C# source visible in the app. The stated audience is Windows developers integrating AI into apps and projects, and the README frames the value in three parts: explore the samples, browse and download models from Hugging Face and GitHub, and export a sample as a standalone Visual Studio project with one click. That last point is what separates it from a documentation site. The repository topics list DirectML, ONNX Runtime, ONNX Runtime GenAI, QNN, NPU, Phi-3, Mistral, Whisper and Stable Diffusion, so the sample set spans text, speech, image generation and hardware-accelerated inference rather than a single modality.

How the gallery, the model store and the exporter fit together

Three mechanisms are visible in the material. The first is a sample catalogue: each entry pairs a task with a model and a UI, and the app runs the sample in process. The second is a model acquisition path. The README states that the app works offline once models are downloaded locally, but you need to be online to pull additional models from Hugging Face or GitHub. So model weights are not vendored in the repository. They land on disk on demand, and the FAQ notes that the app is required to run any sample, which means the catalogue and the downloader are coupled. The third is export. Once a model has been downloaded through the app for a sample, that sample can be exported as a Visual Studio project and run independently of the gallery. The FAQ describes this as the way to try a single sample without the full app. The README also mentions that when executing a sample you can select which model to use, so the sample and the model are not hard-bound to one another. The app also logs basic telemetry, which the privacy statement link covers.

Building and running it from source

The Microsoft Store listing is the shortest path. For a source build, the README gives three steps. Clone the repository with `git clone https://github.com/microsoft/AI-Dev-Gallery.git`, open `AIDevGallery.sln` in Visual Studio, and confirm that the `AIDevGallery` project is the startup project before pressing F5. The prerequisites are Visual Studio 2022 or later with the Windows application development workload, and Windows 10 or newer to run. The README points at the Windows App SDK getting-started page for anyone new to WinUI. There is one architecture trap worth repeating: on ARM64 Copilot+ PCs the solution must be built and run as `ARM64`, not `x64`, and the README says this matters especially for samples that talk to models such as Phi Silica. Device expectations are listed as Windows 10 version 1809 (build 17763) minimum, x64 and ARM64, at least 16 GB of memory recommended, at least 20 GB of free disk recommended, and 8 GB of VRAM recommended for GPU samples. Those are recommendations, not enforced minimums, and the 20 GB figure is a hint about how large the downloaded model files are.

Where AI Dev Gallery stops being the right tool

The gallery is a preview, and the README says so at the top with a request for feedback through GitHub issues. No releases were retrieved for this repository, so there is no tagged version to pin against and no changelog to read before an upgrade. That matters if you were hoping to consume the gallery as a dependency rather than a reference. The FAQ is blunt about the coupling: you need to run the app to try a sample, and only after a model download does export become possible. There is no documented command-line path for pulling a model or running a sample headlessly, and nothing in the material suggests a CI story. The hardware guidance is another boundary. A machine with 8 GB of RAM and integrated graphics will still build the app, but the samples that lean on GPU or NPU execution are the ones the README's VRAM recommendation is aimed at, and the ARM64 note implies that at least some samples silently depend on the build target being right. If your target is a Windows service, a Linux container, or a .NET MAUI app on another platform, the sample code shows a WinUI pattern you will have to rewrite.

AI Dev Gallery against ONNX Runtime GenAI on its own

The closest thing to an alternative is going straight to ONNX Runtime GenAI, which the repository topics name as one of the underlying pieces. The difference is in what each one hands you. ONNX Runtime GenAI is a library: you add a package reference, point it at a model directory, and write the generation loop yourself. AI Dev Gallery is an application that already contains that loop, wrapped in a WinUI shell, plus the model download plumbing and the export step. Choosing the library means you own model acquisition, tokenizer configuration and UI; choosing the gallery means you inherit a working example and then strip out what you do not need. For a developer who has never run a local model, the gallery removes the blank-page problem. For a developer who already knows the runtime and wants to embed inference in an existing WPF or console app, the gallery adds a download of the whole sample application to reach code you could have written from the runtime's own documentation. The gallery is also Windows-only by construction, since it is built on WinUI 3 and the Windows App SDK, while the runtime underneath is not.

Maintenance, upgrade exposure and the MIT licence

The repository is MIT licensed, which is permissive and places few conditions on reuse of the sample code. Two things sit outside that licence and are easy to miss. First, the models themselves. They are downloaded from Hugging Face and GitHub at run time, and each carries its own licence, which the MIT grant on this repository does not cover. Second, trademarks. The README's trademark section states that authorised use of Microsoft trademarks and logos must follow Microsoft's Trademark and Brand Guidelines, and that use in modified versions must not imply Microsoft sponsorship. On maintenance, the material supports a few honest observations. The project is in public preview, contributions require a CLA, and the last push recorded is 2026-09-09, so the codebase is active. But with no releases retrieved, there is no version number to track, which means an update to the gallery can change a sample's behaviour without a changelog entry you can diff. If you export a sample and build a product on it, treat that exported project as your own code from that point: pin your own model files, and do not assume the gallery's next commit will keep the sample compiling.

Who should install it, and what to check first

Install it if you are a Windows developer who writes C# and wants to see a Whisper transcription or a Phi chat loop running locally before you decide how to structure your own app. The preview status and the feedback loop through GitHub issues suggest Microsoft is still shaping the sample set, so early feedback has somewhere to go. Skip it if you need a runtime component with a version number, a headless path for automated testing, or anything that runs outside Windows. Before you build on an exported sample, verify three things in this order: which model file the sample downloaded and where it was written to disk, what licence that specific model carries, and whether the exported project still builds against your installed Windows App SDK version. The ARM64 build-target requirement on Copilot+ PCs is the first thing to confirm if Phi Silica is the sample you came for.

Editorial conclusion

Adopt AI Dev Gallery if you write C# on Windows and want to see how a Whisper, Phi or Stable Diffusion sample is wired up before committing to an architecture. Do not adopt it if you need a headless component for CI, a cross-platform library, or a supported runtime with a versioning contract. Before you build anything on an exported project, check which model file the sample actually pulled, whether that model's own licence permits your use, and whether the sample still compiles against the current Windows App SDK.

Official sources

1. Issues
2. License: MIT
3. microsoft/ai-dev-gallery on GitHub
4. README