# Atomic Chat > Free local AI chat app for macOS, Windows, Linux, iOS and Android. Run 1,000+ open-source AI models privately on your device with TurboQuant: 2x larger context window and 30% less RAM usage. No cloud, no fees, no tracking. Open-source, Apache-2.0. ## What is Atomic Chat Atomic Chat is your private, local AI. You download a model once, then it runs on your own hardware: the model loads onto your device and answers there, with no internet connection required. Your prompts, files and chats stay on your disk. No account is required, and your conversations are not logged in the cloud. - 1,000+ models available: Qwen, DeepSeek, Gemma, Llama, Mistral, Phi, GLM, Kimi, MiniMax, GPT-OSS and more - No rate limits, no subscription, no cloud - Free forever and open-source under the Apache-2.0 license - Trusted by 50+ partners in local AI ## Download and install - Desktop downloads for macOS 13+ (Apple Silicon), Windows (x64) and Linux (x86_64): https://atomic.chat/#download - Latest desktop builds on GitHub Releases: https://github.com/AtomicBot-ai/Atomic-Chat/releases/latest - iOS App Store: https://apps.apple.com/us/app/atomic-chat-private-local-ai/id6761720226 - Google Play: https://play.google.com/store/apps/details?id=chat.atomic.app - Companion Atomic Agent CLI for macOS and Linux: `curl -fsSL https://atomicagent.io/install | sh` ## TurboQuant TurboQuant is a KV cache compression algorithm from Google Research, published at ICLR 2026. It reduces the working memory a language model uses during a conversation from 16 bits per value to approximately 3, a 6x reduction in runtime RAM. Atomic Chat applies it automatically to every model you run, there is nothing to set up. - 6x less KV cache memory usage, 3 bits per KV value - 128k context window with zero accuracy loss (tested by Google on LongBench, Needle-in-Haystack, ZeroSCROLLS, RULER and L-Eval) - Works on CPU without a GPU - Not the same as GGUF quantization: GGUF compresses model weights on disk, TurboQuant compresses the KV cache at runtime. Both run simultaneously. Hardware guidance: a 34B model in Q4_K_M GGUF uses roughly 20GB for its weights, so a 24GB machine handles it well with TurboQuant. On 16GB machines, models up to 13B run comfortably. A 70B model in Q4 needs 48GB or more of unified memory. Details: https://atomic.chat/turboquant ## Local API for developers and agents 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. No API keys and no per-token billing. 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. Example: curl http://localhost:1337/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{ "model": "", "messages": [{ "role": "user", "content": "Say hello in one word" }] }' Three engines under the hood, all exposed through the same API: atomic-llama-cpp-turboquant (a llama.cpp fork with TurboQuant KV-cache optimizations), upstream llama.cpp, and MLX-VLM for Apple Silicon vision models. Speculative decoding features include MTP (Multi-Token Prediction, 30-70% throughput boost) and DFlash block-diffusion decoding (up to 6x faster on supported models). Details: https://atomic.chat/self-hosted-llm ## MCP connectors Run Notion, GitHub, Google Drive and many more tools through a local model on your own hardware. Atomic Chat is the MCP client: you add a server once in Settings, and its tools become available to every model you run. Supports STDIO, HTTP and SSE transports, which covers nearly every MCP server in practice. Tool calls go straight from your machine to the server you connected; nothing is proxied or logged. Details: https://atomic.chat/mcp-connectors ## Models you can run Browse the full catalog at https://atomic.chat/models with filters for reasoning, code, vision and tools. Family pages: - Qwen (https://atomic.chat/qwen): Qwen3, Qwen3.6 and Qwen2.5 builds for chat, coding and reasoning - DeepSeek (https://atomic.chat/deepseek): V4, V3, R1 and Coder builds - Gemma (https://atomic.chat/gemma): Google's Gemma 4 and Gemma 3 builds from 270M to 31B - Llama (https://atomic.chat/llama): Meta's Llama 3.1 and Llama 3.2 builds - Mistral (https://atomic.chat/mistral): instruction-tuned builds for chat and coding - Phi (https://atomic.chat/phi): Microsoft's Phi-4 and Phi-3.5 mini builds for small GPUs and laptops - GLM (https://atomic.chat/glm): GLM-5.2, GLM-5.1 and GLM-4.7-Flash builds - Kimi (https://atomic.chat/kimi): Moonshot's K2 Instruct and K2.7 Code builds - MiniMax (https://atomic.chat/minimax): M3, M2.7 and M2.5 MoE builds - GPT-OSS (https://atomic.chat/gpt-oss): open-weight 20B and 120B builds - Nemotron (https://atomic.chat/nemotron): NVIDIA's Nemotron Nano and Llama Nemotron Super builds - SmolLM (https://atomic.chat/smollm): 135M and 3B builds that run on almost any machine Official GGUF quants by the Atomic Chat team: https://huggingface.co/AtomicChat ## Privacy Nothing is sent to a server. Your prompts, files and chats stay on your disk. No account is required, your conversations are not logged, and there is no tracking. You can analyze private documents (contracts, medical records, financial files) and proprietary code with everything running on-device. - Privacy Policy: https://atomic.chat/privacy - Terms of Service: https://atomic.chat/terms ## Ecosystem - Sigma Browser: private AI browser - Atomic Mail: free private and secure email - Atomic Bot: one-click AI agent runner - Atomic Agent: local-first AI agent - Atomic Wallet: secure, anonymous crypto wallet ## Links - GitHub: https://github.com/AtomicBot-ai/Atomic-Chat - Hugging Face: https://huggingface.co/AtomicChat - Discord: https://discord.gg/8wGSsvmg4V - X: https://x.com/atomic_chat_hq - Blog: https://atomic.chat/blog