A fully ternary Vision Transformer: every weight, including patch embedding, LayerNorm and classifier head, is −1, 0 or +1. It runs right here in your browser. No server, and your image never leaves the page.
Every cell is a real weight from the DeiT-Small checkpoint above: every 48th row and column of each matrix. Cell height is bits per weight, 32 for FP32 and 2 for ternary, so the flattening is the compression. The moving cubes are one token's activations: blue while float, dark while int8 or integer sums. FP32 colours are the latent weights over their absmean scale before rounding. Not shown in full: the position embedding, RMSNorm gains, patch-embedding bias and channel scale stay FP32 ( in total).