| --- |
| title: hyprbench leaderboard |
| emoji: 👻 |
| colorFrom: green |
| colorTo: purple |
| sdk: static |
| pinned: true |
| license: mit |
| short_description: computer-use benchmark for Hyprland - live results |
| --- |
| |
| # hyprbench leaderboard |
|
|
| Live results for [hyprbench](https://github.com/crispwisp/hyprbench) - an |
| OSWorld-style computer-use benchmark for the Hyprland Wayland compositor. |
| 49 tasks across 13 categories (window management, spatial arrangement, |
| browser, canvas pointer-skill, real apps, package installs, terminal-tmux |
| topology, system queries, ...), every verifier reading real state through the |
| system's own introspection channels - never pixels, never exit codes. |
|
|
| Scores are **correctness** (pass rate over scored tasks) and **speed** |
| (end-to-end task latency: median / p90 / max / total). Every run is |
| provenance-stamped: harness CLI at exact version, model, and the hyprbench |
| commit it executed under. |
|
|
| Results are pushed here live by wisp as matrix runs land. |
|
|
