Describe any diagram, Gemma 4 E2B generates it as Excalidraw — entirely in your browser. Desktop Chrome 134+ only.
The LLM outputs compact code (~50 tokens) instead of raw Excalidraw JSON (~5,000 tokens). The TurboQuant algorithm (polar + QJL) compresses the KV cache ~2.4× so longer conversations fit in GPU memory. Needs WebGPU subgroups (Safari/iOS not supported yet) and ~3 GB RAM (mobile browsers cap well below this).
This demo reimplements the TurboQuant algorithm in WGSL compute shaders so it runs on the GPU at 30+ tok/s. The sibling turboquant-wasm npm package implements the same algorithm in WASM+SIMD for CPU-side vector search.
Loading model...