Llama Windows Puts llama.cpp in the Tray and One-Click Hub GGUFs on Windows 11

Hugging Face launched Llama Windows on October 13, 2026: a free open-source WinUI3 app that parks llama.cpp in the Windows 11 notification area and turns Hub GGUF downloads into a few clicks. The announcement post on the Hugging Face blog is dated October 13, 2026. The code lives at ggml-org/Llama-Windows under the MIT License. Release v0.12.0, published October 10, 2026, is the build that adds Hub search from inside the app, richer llama server settings and a set of security hardenings for network-facing use.

On macOS, llama.app already wrapped local GGUF inference for people who did not want to manage binaries by hand. Windows users still had to pick the right llama.cpp build, find a quantized model and launch a server with the right flags. Llama Windows is the Windows 11 counterpart: no terminal required for the default path, and no administrator rights for the llama.cpp installer the app manages.

A tray flyout instead of a taskbar window

The app follows Microsoft Fluent Design and lives in the notification area. Clicking the llama icon opens a compact flyout above the tray rather than a conventional window. Right-click opens or exits. An overlay bound to Alt+Space starts a chat session from anywhere on the desktop. The Available panel lists models already on disk, with load and unload actions and an entry into the llama.cpp web UI. Rows update in place as the managed server changes state.

Recommended Models surfaces ready-to-run quantized Hub builds with parameter count, on-disk size and whether vision is supported, so a user can judge fit before downloading. One click starts the transfer with live progress. Models land in the standard Hugging Face cache under %USERPROFILE%\.cache\huggingface\hub, shared with other Hub tooling, so a file downloaded once is reusable outside the tray app.

A local server builders can call

On launch, Llama Windows attaches to a llama.cpp server that is already running or starts the one it manages. If llama.cpp is missing, it installs it with the official installer. The managed server listens at http://localhost:9931 by default and speaks an OpenAI-compatible API. Scripts and agents can list models with a simple GET to /models and point any OpenAI client at http://localhost:9931/v1 without leaving the machine.

That is the builder-facing half of the product. The tray UI is for discovery and load/unload. The same process is a local endpoint for Hermes-style agents, evaluation scripts or IDE plugins that already speak OpenAI wire format. Settings accept a Hugging Face token for gated or private repos, and the model cache can be moved to another drive. Daily logs are kept for seven days under %LOCALAPPDATA%\LlamaApp\logs.

Release v0.12.0 tightens the network story further. An API key protects the server when it is reachable beyond the local machine. The Hugging Face token is stored encrypted. The installer download is verified and size-bounded. The embedded chat overlay is constrained so it cannot navigate away from the local server. Operators can choose the listen interface, cap how many models run in parallel, quantize the KV cache and pass extra llama serve flags.

Windows on Arm and RTX Spark on day one

The launch post states native support for Windows on Arm and for NVIDIA’s RTX Spark, with llama.cpp using the Spark GPU without manual tuning. The framing is on-device privacy: prompts and generations stay on the PC unless the user points a client somewhere else. Hugging Face says it worked with NVIDIA so the tray app was ready for RTX Spark at announcement.

The GitHub repository was created in June 2026 and has been pushed through the October 13 announcement window; star count is still modest compared with llama.cpp itself. That is expected for a thin WinUI shell. The substance is distribution: GGUF remains the common local format on the Hub, and Windows 11 users finally get a first-party-adjacent path that does not start with a compiler toolchain.

For Local AI builders, Llama Windows is less a new inference engine than a packaging decision. llama.cpp stays the runtime. The Hub stays the model store. The tray app removes the friction that kept Windows setups one terminal command away from “it works on my Mac.” Anyone already comfortable with CLI servers can ignore the flyout and keep calling port 9931. Anyone who was blocked by that CLI now has a MIT-licensed on-ramp that still exposes the same OpenAI-compatible socket underneath.