Browser-Based Local LLM Chat
Runs on your GPU · chats saved on this device
How it works & FAQ ↓Start a Conversation
Chat with AI running entirely in your browser. Your chats are saved only on this device.
Qwen3.5-0.8B isn't on this device yet
It needs a one-time 0.57 GB download. After that it's cached and starts in a few seconds. Replies are generated on this device; only tool lookups, like weather or Wikipedia, go online.
Try one of these prompts:
About Browser-Based LLM Chat
Experience local multimodal & reasoning LLM conversations running entirely in your browser: the model runs on your device and your chat history stays in your browser's storage. Powered by Transformers.js, high-performance in-browser LLM inference using WebGPU.
How to Use
- Select a Model: Choose from available models like Qwen3.5-0.8B, Qwen3.5-2B, LFM3-VL-2B, Ternary Bonsai 2 4B Vision, Gemma 4 E4B Mobile Vision, and Qwen3-8B based on your needs and hardware.
- Wait for Download: On first use, the model will download and cache in your browser for future sessions.
- Start Chatting: Type your message and press Enter. Responses stream in real-time. A message sent while the model is still loading goes out as soon as it is ready.
- AI Tools & MCP: Switch real-time tools (web search, web scraping) on or off with the Tools button under the message box, or connect your own MCP servers via Streamable HTTP in Settings › Connections.
Browser Support
WebGPU is required for GPU-accelerated inference. Chrome and Edge have the best support. A dedicated GPU with 2-8GB VRAM is recommended for optimal performance.
Security & Privacy
Private by design: The model runs locally in your browser, so your prompts and replies are never sent to a chat server. Chat history is stored only in your browser's local storage and can be cleared anytime.
Tools are the exception: when the model uses one, that tool's input (for example a city name or a search term) is sent to the service it calls: Open-Meteo for weather, CodeSamplez for web search, Wikipedia for lookups, or any MCP server you add. Turn tools off in the Tools menu to keep everything on this device. Model weights are downloaded once from Hugging Face and cached in your browser.
Frequently Asked Questions
On first use, the selected model downloads and is cached in your browser. Download size and load time vary by model. Larger models need more VRAM and take longer to load; cached models usually load faster on later visits.
WebGPU is a new web standard that may not be enabled by default in all browsers. In Chrome/Edge, go to chrome://flags and enable "WebGPU". In Firefox, set "dom.webgpu.enabled" to true in about:config. Also ensure your GPU drivers are up to date.
Qwen3.5-0.8B is the default for most devices. LFM3-VL-2B and Ternary Bonsai 2 4B Vision and Gemma 4 E4B Mobile Vision add image input. Available models: Qwen3.5-0.8B, Qwen3.5-2B, LFM3-VL-2B, Ternary Bonsai 2 4B Vision, Gemma 4 E4B Mobile Vision, and Qwen3-8B.
Yes! When using multimodal models such as LFM3-VL-2B or Ternary Bonsai 2 4B Vision or Gemma 4 E4B Mobile Vision, you can attach an image via the image button or paste an image directly from your clipboard into the input field.
Voice Mode runs 100% locally on your device. Speech-to-text dictation uses Whisper running inside a browser Web Worker via WebAssembly, so your audio is never uploaded to any server. Text-to-speech reading aloud uses your browser's native SpeechSynthesis API completely offline.
Only in your browser. Replies are generated on your device with WebGPU, and your conversations are saved in your browser's IndexedDB storage, never on a server. You can export or delete them at any time. If tools are on and the model uses one, only that tool's input (such as a city or search term) is sent to the tool's service: Open-Meteo for weather, CodeSamplez for web search, Wikipedia for lookups, or an MCP server you added.
You might also like
Explore more interactive tools and demos.
