Atomic Chat

Local AI app and inference engine for agents. Run open-weight LLMs locally — private, on your machine.

Getting Started · Hugging Face · Discord · X / Twitter · Bug Reports

📦 Download

Desktop

Linux packages — from the Atomic Chat repository, updated by your package manager (how to install)

Mobile

🔌 Use It as an API

Atomic Chat runs an OpenAI-compatible server at http://localhost:1337/v1 — a drop-in replacement for the OpenAI SDK. Load a model in the app, then point any client at it:

curl http://localhost:1337/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "<model-id-loaded-in-atomic-chat>",
    "messages": [{ "role": "user", "content": "Say hello in one word" }]
  }'

from openai import OpenAI

# Atomic Chat is OpenAI API-compatible — only the base_url changes.
client = OpenAI(base_url="http://localhost:1337/v1", api_key="not-needed")

resp = client.chat.completions.create(
    model="<model-id-loaded-in-atomic-chat>",
    messages=[{"role": "user", "content": "Say hello in one word"}],
)
print(resp.choices[0].message.content)

Bound to 127.0.0.1 by default; set host: 0.0.0.0 to expose it on your LAN. Works with any agent, CLI, or IDE plugin that speaks the OpenAI API — see Launch With below.

✨ Features

Local models

• Run open-weight LLMs locally from HuggingFace — Llama, Gemma, Qwen, Mistral, Phi, and others
• Multi-Token Prediction (MTP) speculative decoding — 30–70% throughput boost on supported models, up to 3× on Gemma 4
• DFlash block-diffusion decoding — up to 6× faster on Qwen 3.6, Gemma 4, Kimi K2.5
• Flash Attention toggle (on / off / auto)
• Automatic reasoning-context tracking for chain-of-thought models
• Auto context-window expansion with overflow notifications
• EAGLE-3 speculative decoding for Gemma 4 on Apple Silicon (MLX)
• MTP on MLX for Qwen 3.5 / 3.6 and DeepSeek V4
• TurboQuant KV cache (turbo3 / turbo4) on llama.cpp — now on Windows & Linux too, not just macOS: up to ~4.3× smaller KV cache footprint, CPU and GPU (CUDA / Vulkan)
• TurboQuant KV cache on MLX-VLM — smaller memory footprint via RHT-correct fast paths

Cloud models

• Built-in providers: OpenAI, Anthropic, Mistral, Groq, MiniMax, Qwen, Moonshot
• Bring your own key, switch model per chat, mix local and cloud freely

Tools & integrations

• One-click agent launch — launch agents like Atomic Agent, Claude Code, Codex CLI, Cline, OpenCode, Droid, Goose, OpenHands, Copilot CLI, Kilo Code, Zed and ZCode in one click from the Integrations tab
• Artifacts — live preview panel for HTML/CSS/JS code with copy, download and print
• Connect multiple MCP servers — bring your own tools, file access, web search
• Custom assistants with per-assistant system prompts
• Projects with conversation tree view in the sidebar

Local API

• OpenAI-compatible server at http://localhost:1337/v1 — drop-in replacement for the OpenAI SDK
• Works with any agent, CLI, or IDE plugin that speaks the OpenAI API
• Bound to 127.0.0.1 by default; set host: 0.0.0.0 to expose on LAN

Privacy

• Everything runs locally when you want it to — local server is loopback-only by default
• Your conversations and keys stay on your machine

⚙️ Inference Engines

Three engines under the hood, all exposed through one OpenAI-compatible API at http://localhost:1337/v1:

• atomic-llama-cpp-turboquant — our llama.cpp fork with TurboQuant KV-cache optimizations (turbo3 / turbo4) for faster, lower-memory quantized inference. Now a selectable second provider ("Atomic Llama.cpp Turboquant") on all three desktops — macOS, Windows, and Linux — CPU and GPU (CUDA / Vulkan).
• Upstream llama.cpp — official ggml-org build, the default engine on Windows and Linux for the widest hardware coverage and MTP support.
• MLX-VLM — Apple Silicon-native engine for vision-language models, running on the Neural Engine and unified memory. Faster than llama.cpp on M-series chips for supported models.

Speculative-decoding features available across backends:

• MTP (Multi-Token Prediction) — a draft model predicts ahead, the full model verifies in one pass. Available on macOS and Windows.
• DFlash — block-diffusion speculative decoding for Qwen 3.6, Gemma 4, Kimi K2.5 and others. Apple Silicon only; can't be enabled together with MTP.
• Flash Attention — Settings → on / off / auto.

Tools talking to http://localhost:1337/v1 don't need to know which backend is running underneath — switch engines without reconfiguring clients.

The engines are driven by atomic-chat-core — the core of Atomic Chat, in its own repository. It doesn't do inference itself; it manages the engines: downloads and launches backends, manages models, routes requests between local engines and cloud providers, and serves the OpenAI-compatible API on :1337. The desktop app ships a pinned core release; the core also runs on its own as a CLI (atomic-chat-core serve …) or as a library.

🚀 Launch With

Atomic Chat runs an OpenAI-compatible server at http://localhost:1337/v1, so any agent, CLI, IDE plugin, or app that speaks the OpenAI API can run on top of your local models — no extra glue needed. Just point its base URL at Atomic Chat and you're done.

A few projects already ship first-class support with their own setup docs: