W4M · Research Notes

Blog

Results from our Graph of Models work. What it does — not how.

June 2026

99.82% of 13,000 hard Sudoku puzzles. 97,000 parameters.

A 13,000-puzzle hard benchmark, solved at 120ms per puzzle by a model five orders of magnitude smaller than a frontier LLM — trained in 26 minutes on a MacBook. With an interactive demo.

Read + watch it solve →