A 27M-param model that solves hard Sudoku/mazes where LLMs fail, without CoTgithub.com/sapientinc10 pointsmingli_yuana year ago