TiinyBench Measures the Box, Not the Spec Sheet

Field notes

The benchmark runs on the Tiiny Pocket itself and separates measurements by model class instead of forcing text, speech, image, and embedding workloads onto one ranking axis. Its reported tests include four prompt lengths from 72 to 6,260 tokens, a 1,500-token sustained generation, and concurrency runs from one to eight requests. The page also documents a measurement error in its first utilisation method and the revised sampling approach.