Self-host Atomic Chat

Atomic Chat is a cross-platform desktop and mobile application for running large language models on personal hardware. It provides a graphical chat interface and runs a local API server compatible with the OpenAI SDK at localhost:1337/v1, allowing any compatible client to connect. The app supports over a thousand models from Hugging Face including Llama, Qwen, DeepSeek, Gemma, and Mistral in GGUF, MLX, and ONNX formats. Three inference engines are available: a custom llama.cpp fork with TurboQuant KV-cache optimization, upstream llama.cpp, and MLX-VLM for Apple Silicon. Features include Multi-Token Prediction and DFlash speculative decoding, Flash Attention, MCP server connections for tools and file access, artifacts for live HTML/CSS/JS preview, custom assistants with system prompts, project organization with conversation trees, and one-click launching of AI agents like OpenCode, Goose, Cline, Kilo Code, and others. The app is built with Tauri and Node.js, distributed as native installers for macOS (universal DMG), Windows (x64/ARM64 exe), Linux (AppImage), iOS (App Store), and Android (Google Play). System requirements include macOS 13.6+, Windows 10/11 x64, or Linux x86_64 with glibc 2.35+, with memory needs of roughly 8GB for smaller models, 16GB for medium ones, and 32GB for larger models. The project is licensed under Apache 2.0 and originated as a fork of Jan by Menlo Research.