What's this about?
Alongside K3, Moonshot released PerceptionBench, a benchmark that separates raw visual perception from reasoning. Instead of inventing test categories up front, they studied how today's frontier models fail across 42 existing benchmarks, distilled 10 "atomic" perceptual skills from those failures, and wrote 3,000 verified questions — each testing one skill and answerable purely by looking, with no reasoning or outside knowledge needed.
What exactly is "visual perception" vs. reasoning?
Perception is seeing what's there — colors, counts, positions, shapes. Reasoning is thinking about it — inferring, calculating, relating. Most vision benchmarks tangle the two, so a model can flub a question either by mis-seeing or mis-thinking, and you can't tell which. PerceptionBench isolates the seeing, so a wrong answer means the eyes failed, not the brain.
Why does it matter?
If you can't measure perception cleanly, you can't fix it. As models take actions in the physical and on-screen world, "did it actually see the thing" becomes a load-bearing question — and a benchmark derived from real failure modes is harder to game than one invented in advance.