Mano-P agent plays Chinese mahjong via pure vision

// 9h agoVIDEO

Mano-P agent plays Chinese mahjong via pure vision

Mininglamp AI showcased its open-source Mano-P GUI-VLA agent playing Chinese Mahjong entirely through screen vision and mouse clicks. The demonstration serves as a brutal stress test for the model's ability to operate in complex, unstructured visual environments without underlying APIs.

// ANALYSIS

Testing a GUI agent on Mahjong is a brilliant flex that proves visual-action models are graduating past predictable web DOMs into messy, unstructured visual spaces.

–Mano-P relies on raw pixel perception, making decisions based purely on the screen state without any backend hooks or game data
–The game demands high visual precision to distinguish intricate tiles and fast reasoning to react to opponent actions
–Unlike cloud-dependent models, Mano-P is heavily optimized to run locally on consumer edge hardware like M4 Mac minis
–The project currently tops the OSWorld benchmark for specialized GUI models, offering a compelling open-source alternative for computer-use tasks

// TAGS

mano-pvisionmultimodalcomputer-useagentopen-sourcelocal-firstedge-ai

DISCOVERED

9h ago

2026-05-28

PUBLISHED

12h ago

2026-05-28

RELEVANCE

8/ 10

AUTHOR

Enough-Astronaut9278