ggml-org / Llama-Windows

# Llama
**Local LLMs, one click away — right from your Windows system tray.**
Llama is a Windows 11 tray app for running llama.cpp models locally. It's a WinUI 3 port of Llama for Mac.
## Install
Grab the latest `.msixbundle` (x64 + ARM64) from **Releases** and double-click to install.
## Highlights
- **Lives in your tray** — a borderless, Mica-backed flyout anchored to the tray icon
- **Zero-setup llama.cpp** — uses your existing `llama.exe`, or downloads one for you
- **One-click models** — browse and download recommended models from the Hugging Face Hub
- **Model details** — click any model for a context-length picker with per-option memory estimates (read from the GGUF header), remembered per model, plus chat, copy-model-ID, curl-request, and Hugging Face actions
- **Standard storage** — models live in the Hugging Face cache, shared with `llama.cpp` and other tools
- **Spotlight-like overlay** — press `Alt+Space` from any app to chat with your loaded model
## The overlay
`Alt+Space` from any app opens a borderless window with the WebUI for whichever model is loaded. It hides instead of closing, so re-summoning is instant and your conversation is still there.
## How it works
Llama manages a `llama serve` process and talks to it over its REST API. Your models stay in the standard Hugging Face cache (`%USERPROFILE%\.cache\huggingface\hub`), shared with `llama.cpp` and HF tooling. Settings and logs live under `%LOCALAPPDATA%\Llama`.
## Privacy
Llama sends no telemetry — no events, no analytics, no install id. The only thing outbound HTTP requests (model downloads, release-update checks, and calls to the local llama-server) carry is the standard User-Agent header:
 [code] 
— the app version, the Windows build, and the processor architecture. Prompts and inference stay 100% local.
## Building from source
You'll need the **.NET 10 SDK** on Windows (ARM64 or x64 host; the SDK cross-builds both).
The WinUI app is built per-platform — pass the matching `-p:Platform` alongside the `-r` runtime identifier:
 [code] bash
dotnet publish LlamaApp.csproj -c Release -r win-arm64 -p:Platform=ARM64 \
 --self-contained true -p:PublishReadyToRun=false -p:PublishSingleFile=false \
 -o publish/win-arm64
 [code] bash
dotnet test LlamaApp.Tests/LlamaApp.Tests.csproj -c Release --arch x64
```
`LlamaApp.LlamaCpp.Tests` is platform-agnostic and runs natively.
### Building the MSIX
The installed release is a signed `.msix`, produced by the CI
(`.github/workflows/_build.yml`) with Visual
Studio MSBuild and the `GenerateAppxPackage` target — `dotnet build` alone
does not pack the Appx. `dotnet publish` above is sufficient for local use.
---
Made with 🦙 for Windows 11 · Models run 100% locally — your prompts never leave your machine.