Free app lets Windows users run AI models on their own PC

A new app, Llama Windows, aims to make it easier to run AI language models directly on your own computer. The app is free and open source, and it places the llama.cpp tool in the Windows 11 system tray. That lets users download and run open language models locally without opening a terminal, the text-based command window.

The information here is based on the developers' own blog on Hugging Face, dated October 13, 2026, and on the project's page on GitHub. The features are therefore the maker's own descriptions and have not been independently tested.

How it works

Models are downloaded with one click from the Hugging Face Hub, and download progress is shown along the way. A panel of recommended models shows popular quantized versions. Quantization means a model has been made smaller and easier to run on ordinary hardware. The panel also shows the number of parameters, disk size and whether the model can handle images. The app installs and manages the llama.cpp runtime itself.

A chat overlay can be opened from any app with Alt+Space. There is also a context length selector, which sets how much text the model can keep in mind at a time, with an estimate of memory usage. The app can also use gated or private models via an access token.

For developers, the app runs a local server at http://localhost:9931 that is compatible with OpenAI's API. That means other programs can talk to the local model the same way they would talk to an online service.

Requirements and installation

The app requires Windows 11 but not administrator rights. It supports Windows on Arm and acceleration with NVIDIA graphics cards. Builds are available for x64, ARM64 and x86. It is installed by downloading an .msixbundle file from the project's Releases on GitHub or via llama.app.

Models are stored in Hugging Face's shared cache folder. Settings and logs are located under %LOCALAPPDATA%\Llama, and daily logs are kept for 7 days.

Privacy and license

According to the project, the app sends no telemetry. Only a standard User-Agent with the app version, Windows build and processor architecture is sent. The project is MIT-licensed. At the time of research, the repository had 135 stars, 18 forks and 6 open issues.

The app is a WinUI 3 app and a Windows counterpart to an existing Llama app for Mac. The sources do not make it clear whether ggml-org or Hugging Face is formally the publisher: The code is hosted by ggml-org, while the blog is on Hugging Face.