Overview

Llama is a single‑file system‑tray application that runs open‑source language models directly on a Windows PC. The program presents a tiny menu‑bar icon and offers two main functions: a web‑based chat window and a local API that mimics the OpenAI REST endpoints.

When you launch Llama, a small window appears containing “Open chat” and “Open API” buttons. Choosing “Open chat” opens a browser tab pointing to `http://localhost:9931`. The interface displays the selected model's name and allows you to type free‑form prompts. The model processes the input and returns a JSON‑encoded response that the UI renders as plain text. Typical inputs include questions, code snippets, or document excerpts, and the outputs are tokenized completions formatted like standard ChatGPT replies.

Selecting “Open API” turns the process into a local service listening on port 9931. Any application that sends an HTTP POST to `/v1/chat/completions`, with headers `Content-Type: application/json` and a body containing `{"model": ".", "messages": [{"role": "user", "content": "."}]}`, receives a response that follows the OpenAI specification. This means code editors, IDE extensions, or custom scripts that already talk to OpenAI can point to the local endpoint without changing their logic. The service supports streaming, tool calling, structured outputs, and vision endpoints if the underlying model enables them.

The software is designed for personal or small‑team computing. It is ideally suited for developers or writers who want full control over the model and data privacy. People with limited storage or who rely on cloud‑based APIs to avoid training or downloading models might find Llama less convenient. The application is free, open source, and has no configuration step beyond accepting the default installation; however, the listing does not provide any pricing tiers, support plans, or licensing details beyond the MIT/Apache‑like license implied by the open‑source status.

Installation is just a 1 MB download for Windows 11. After extraction, the app runs with no further setup. Models are fetched from the Hugging Face cache and stored once, regardless of how many programs use them. When idle, the service stops loading models after five minutes, conserving RAM and disk space. Model recommendation is automatic: the first time you open the app, Llama scans your PC's memory and suggests a default model that fits within your resources, such as Gemma‑4 12B for a 32 GB system.