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

Not an LLM, but the ternary findings should carry over, and we hadn't seen Sherry-style 3:4 weights run in a browser before. Disclosure: this is our work at Precisit, everything is MIT.

What it is

• A 7.4M-parameter one-pass scorer (the jevlike family): the board goes in, one score per legal column comes out. No search.
• Weights in T34, Sherry's 3:4 format: in every four weights one is zero and three are ±1, so four weights fit in 5 bits. One fp16 scale per 128 weights gives 1.375 bits per weight. The embedding is int8; norms and biases are fp16.
• It runs in the browser on a small WebGPU runtime: 1.1 ms per move (idle M5 Pro, Chrome).

200 games each, both sides play a random move 5% of the time, a win counts 1 and a draw ½.

What we learned

1. Converting the finished model to 3:4 collapsed it (0.13 against the depth-4 bot). Training with the format in the forward pass fixed it completely, whether from scratch or fine-tuning.
2. Attention's q/k/v matrices are the sensitive ones. Group size (64/128/256) barely mattered.
3. Seeds matter: two runs of the same T34 recipe scored 0.945 and 0.882.

Play it:
https://precisit.github.io/onepass-web/demo/c4-size/

Code, models, every result:
https://github.com/precisit/onepass-webgpu-ternary

The write-up:
https://precisit.com/en/blog/onepass-c4-size/

Has anyone gotten post-training 3:4 conversion to work on models, or does it need training?