Benchmarks

OSWorld 3.0

Agentsunit % · normalized over [0, 82]

OSWorld 3.0 — third-generation real-desktop-GUI computer-use benchmark with longer-horizon multi-app tasks, scored pass@1. Distinct from and not comparable to OSWorld 2.0 or OSWorld-Verified (v1).

#ModelSourceScoreNormalized

Score distribution

2 tracked results across the normalization window

082

Score vs. parameters

Open-weights models, log-x params

Only one open-weights model with a disclosed parameter count — not enough to plot a trend.