SkillsBench with Skills - 86 tasks, weighted test reward
Agents · 2026-02-13
The 86-task SkillsBench set as run in the Muse Glimmer report: each task runs in its own container with the relevant skill folders mounted and bash, image-view and browser tools. A hidden pytest suite grades the result, and the task reward is the weighted fraction of tests passed. The task count and partial-credit reward differ from the 87-task v1.1 leaderboard's resolution rate, so this is a diagnostic.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.6 27B | 46.6 |
| Muse Glimmer 30B | 44.3 |
| Gemma 4 31B IT | 32.4 |