ThoxyWeb. Chat with THOXY in your browser.

THOXY is ThoxyWeb's in-browser assistant — Gemma 4 E2B (QAT Mobile) running fully on-device with WebGPU. Weights cache locally after the first load, and nothing you type ever leaves your machine.

2.3BEffective params

128KContext window

~250tok/s · M4 Max

100%On-device

Runs fully on-device — nothing leaves your machine

Kernels

What are Kernels?

Kernels are the low-level GPU programs that do the model's actual math — the matrix multiplications, attention, and normalization behind every token. And how well they're optimized can dramatically speed up inference.

WebGPU & WGSL. Each kernel is a WebGPU compute shader, written in WGSL — the language that runs general-purpose math on the GPU — entirely locally in your browser.
Agentic Kernel Optimization. Every kernel was generated by AI (originally by Fable 5, before it was shut down; preserved as-is in this THOX build), benchmarked on an Apple M4 Max, and refined through an evolutionary, genetic-style search toward the fastest version.
Blazingly Fast. ThoxyWeb runs Gemma 4 E2B at up to ~250 tokens/sec on an M4-class GPU, pushing your device to its limits.

Select a kernel to read its real source