Atomic Chat offers open-source alternative that runs models locally with OpenAI-compatible API

Atomic Chat defines itself as an open-source alternative to ChatGPT that runs one hundred percent offline on your computer. The application, available for desktop and mobile, runs open-weight models from Hugging Face such as Llama, Gemma, Qwen, Mistral and Phi, and exposes a local server compatible with the OpenAI API at 127.0.0.1:1234. Any client that speaks the OpenAI SDK, from a simple curl to IDE plugins, can point its base_url to the local server and start working without an API key and without network traffic.

Three inference engines operate behind the same interface. The first, atomic-llama-cpp-turboquant, is a fork of llama.cpp with TurboQuant KV-cache optimization, turbo3 and turbo4 variants, reducing memory footprint by up to 4.3 × and now available on Windows and Linux (CPU and GPU via CUDA/Vulkan), not only on macOS. The second is the official ggml-org llama.cpp, used as the default on Windows and Linux because of broad hardware coverage and support for Multi-Token Prediction (MTP). The third, MLX-VLM, leverages Apple’s Neural Engine and the unified memory of M- chips for vision-language models, surpassing llama.cpp performance on supported models.

Speculative decoding features are available on all three engines. Multi-Token Prediction (MTP), a draft model that predicts several tokens at once, provides a 30-70 % throughput boost on supported models and reaches up to three-fold on Gemma 4. DFlash block-diffusion decoding pushes speed up to six-fold on Qwen 3.6, Gemma 4 and K2.5. On Apple silicon, EAGLE-3 adds acceleration to Gemma 4 via MLX, and MTP on MLX is active for Qwen 3.5/3.6 and DeepSeek V4. A Flash Attention toggle (on/off/auto) allows manual control when needed.

Cloud access is provided when required, but the local server is the default. Users who need closed models receive built-in access to cloud providers OpenAI, Anthropic, Mistral, Groq, MiniMax, Qwen and Moonshot, with a BYOK personal key and the ability to swap models in any chat or mix local and cloud in the same project. The local server listens only on loopback by default; setting the host to 0.0.0.0 exposes it on the local network for agents running from another machine in the LAN.

The integrations tab enables one-click launch of popular agents: Atomic Agent, Claude Code, Codex CLI, Cline, OpenCode, Droid, Goose, OpenHands, Copilot CLI, Kilo Code and Zed. An Artifacts panel shows a live view of HTML/CSS/JS code with copy, download and print options. Support for MCP (Model Context Protocol) lets you connect multiple tool servers, access files, perform web searches and more. Custom assistants with system prompts and a conversation-tree view in the side bar complete a full workspace without leaving the app.

Conversations and keys remain on the machine. No hidden telemetry, no automatic upload; as long as a local model is chosen, everything runs inside the user’s sandbox. For developers building internal AI tools, researchers testing models without sending data outward, or anyone who simply wants a chatbot that works offline, Atomic Chat provides the infrastructure without requiring a cloud account.