Browser AI Playground
Run Bonsai 27B & Gemma 4 Vision in Your Browser
WebGPU is unlocking local inference at absurd scale. A 27B parameter model running directly in Chrome, on a 4GB Android phone — no server, no API, no cost. Just shaders, attention kernels, and the browser.
Requires WebGPU-compatible browser. Chrome 113+, Edge 113+, or Chrome Canary. On some setups you may need to enable chrome://flags/#enable-unsafe-webgpu. Firefox support is experimental. Safari 18+ has partial WebGPU support. 4GB+ RAM recommended for 27B model (quantized to ~3-4 bits).
🧠 How does Bonsai 27B run in the browser?
WebML community compiles the model into WGSL compute shaders. Weights are quantized to 3-4 bits, KV-cache is tiled across GPU memory. The entire inference loop — tokenization, attention, FFN, sampling — happens on your GPU via WebGPU compute passes. Nothing leaves your device.
👁️ Gemma 4 Vision in-browser
Google's Gemma team ships a vision-language model that can caption images, answer visual questions, and reason over what it sees — all via WebGPU. Upload an image, ask a question, watch it think. No upload to a cloud.
Bonsai 27B — WebGPU Chat
27B param model running locally via WebGPU shaders. Try long-form reasoning, coding, creative writing.
Loading Bonsai 27B — WebGPU Chat...
Gemma 4 Vision — Interactive (via HF Space)
Vision-language model from @googlegemma tweet. Upload images + Q&A. Interactive Space — open in HF for full demo.
Loading Gemma 4 Vision — Interactive (via HF Space)...
Spaces hosted on HuggingFace. If an embed fails to load, use the "Open in HF" button. Some models may take 30–60s to initialize as weights download and shaders compile. First token is slowest; subsequent tokens stream fast. Built with WebGPU + WGSL.