# Atomic Chat

A local desktop and mobile app that runs open-weight LLMs offline and exposes them through a drop-in, OpenAI-compatible local API for agents, CLIs and IDE plugins.

Atomic Chat is a local AI app and inference engine for running open-weight language models such as Llama, Gemma, Qwen, Mistral and Phi entirely on a user's own machine, with desktop and mobile builds. Loading a model in the app starts an OpenAI-compatible server bound to localhost by default, so any client that speaks the OpenAI API, including agents, CLIs and IDE plugins, can point at it with only the base URL changed; setting the host to 0.0.0.0 exposes it to the local network.

It supports speculative decoding techniques such as multi-token prediction and block-diffusion decoding for a throughput boost on supported models, a key-value cache quantization mode that shrinks memory footprint on CPU and GPU across Windows, Linux and macOS, and built-in cloud providers including OpenAI, Anthropic, Mistral, Groq, MiniMax, Qwen and Moonshot for mixing local and cloud models in the same chat. It can launch other coding agents, such as Claude Code, Codex CLI and OpenCode, directly from an integrations tab, and includes a live preview panel for generated HTML, CSS and JavaScript. It is written in TypeScript.

Website: https://atomic.chat