Does anything I type get sent to a server?
No. The model runs on your GPU inside the browser. The only network traffic is the one-time download of the model files from Hugging Face and the model library from GitHub.
Which browsers work?
Any browser with WebGPU enabled: current Chrome and Edge on desktop and Android, Safari on recent macOS and iOS, and Firefox releases that ship WebGPU. The tool detects support at runtime and explains what is missing if it cannot run.
How much space and memory do I need?
The smallest models are around 200–400 MB to download and need under 2 GB of GPU memory. Mid-size models are 1–2.5 GB and suit 8 GB machines. The catalog shows the exact download size and GPU-memory need for every model and marks whether it fits your device.
Can I use it offline?
Yes. After a model has been downloaded once, the browser keeps it and the chat works without a connection.
How do I add my own model?
Open Models, choose Add model, and paste the URL of a repository with MLC-compiled weights (mlc-chat-config.json plus tensor-cache.json) together with the WebGPU model library built for that model. The tool fetches the files to check them before it saves anything.
What is the Reasoning block?
Some models (Qwen3, Qwen3.5, DeepSeek R1) think before they answer. That thinking streams into a collapsible block above the answer and is never sent back to the model as part of the conversation.
Why is a model marked as not fitting?
The verdict compares the model's GPU-memory need with your device's memory, checks for required GPU features such as 16-bit shaders, and looks at free storage for the download. You can still try a model; the tool will tell you if the GPU runs out of memory.
Where are my chats stored, and how do I delete them?
In this browser's IndexedDB. Delete a chat from its menu, or everything at once from Settings. Clearing the site's data in the browser also removes them.
Can I stop a long answer or cancel a download?
Yes. Stop replaces the send button while an answer streams, and every download has a Cancel; finished parts stay cached, so a retry resumes where it left off.
