
An open-source benchmark of how well LLMs solve nonogram puzzles, from 5x5 to 20x20.
An open-source benchmark of how well LLMs solve nonogram puzzles, from 5x5 to 20x20. Query every result through a REST API or MCP server.
https://www.nonobench.com/mcpNonobench measures how well LLMs solve nonogram (picross) puzzles: 40 puzzles across 5x5, 10x10, 15x15, 20x20 grids. Standard is 5x5, 10x10 and 15x15, and overall accuracy covers those 30 puzzles. Hard mode is ten 20x20 puzzles, reported only per size (size "20x20"). Accuracy is the share of puzzles where the model's grid satisfies every row and column clue. Results last updated 2026-10-26T19:22:58.543Z.
None in the store listings or on the server