Sherry's 3:4 ternary format (1.375 bits per weight) running on WebGPU: a 1.6 MB model that plays Connect Four as well as its 7.8 MB int8 version

A 1.6 MB Connect Four model, using Sherry's 3:4 ternary format (1.375 bits per weight), is now running on WebGPU. This model performs as well as its 7.8 MB int8 version, demonstrating that ternary findings could extend beyond this specific application. The T34, trained ternary model, achieved a 0.93 score against a depth-4 bot and 0.91 against a depth-6 bot, outperforming other configurations in a 200-game test.

This is the first time Sherry-style 3:4 weights have been shown running in a browser, unlike previous demonstrations that did not specify the execution environment.