Bonsai 1-bit WebGPU

A browser demo that runs a 1-bit quantized LLM locally using WebGPU, meaning zero server calls and zero API keys. The 1-bit angle is the actual story here: at this level of quantization, weights are stored as single bits rather than 16- or 32-bit floats, which collapses memory usage dramatically and makes it plausible to run inference on consumer hardware — or in a browser tab. The practical quality ceiling is noticeably lower than a full-precision model, so this is not something you would ship to users today. Where it gets interesting is as a reference point: if WebGPU inference at 1-bit is this far along, the gap to useful on-device models is closing faster than most roadmaps suggest. Worth an afternoon if you are building anything where data privacy or offline operation is a hard requirement and you want to understand the current floor. Reservation: thin on documentation and model-capability details. -> Best for: AI engineer or indie hacker exploring on-device inference constraints