A small model fills in the whole grid at once, then corrects it over 12 passes. It runs on your device. How it works
Most language models write one token at a time. A Next-Concept Model drafts the whole answer at once, then corrects it. Here it fills all 81 cells in the first pass, mostly wrong, and fixes them over the next 11. Watch the red cells disappear.
The model runs once and returns all 12 passes, then the page plays them back. It runs in your browser, so your puzzle never leaves your device. Each difficulty is a separate model of about 150 MB, downloaded once.
Based on Kumbong et al., 2026 (code). Models exported to ONNX by mrfakename. Both MIT. Source: github.com/yrangana/ncm-lab.