Desktop Task Completion on OSWorld
60.1AccuracyKimi K2.5
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Kimi K2.5Model Type=Standalone Large2026.04 | 60.1 | 8.3 | 0.132 | 22.4 | — | — | — | |
| Qwen3-VL-8B + Kimi K2.5Small Model=Qwen3-VL-8B, Large Model=Kimi K2.5, Mode=Step-level Cascading2026.04 | 59.3 | 6.5 | 0.078 | 23.7 | 240 | 37.4 | 62.6 | |
| EvoCUA-8B + Kimi K2.5Small Model=EvoCUA-8B, Large Model=Kimi K2.5, Mode=Step-level Cascading2026.04 | 58.2 | 4.5 | 0.051 | 25.2 | 173 | 59.5 | 40.5 | |
| Claude Sonnet 4.5Model Type=Standalone Large2026.04 | 58.1 | 6.4 | 0.881 | 25.4 | — | — | — | |
| EvoCUA-8B + Claude Sonnet 4.5Small Model=EvoCUA-8B, Large Model=Claude Sonnet 4.5, Mode=Step-level Cascading2026.04 | 55.4 | 4.1 | 0.224 | 26.2 | 168 | 60.6 | 39.4 | |
| Qwen3-VL-8B + Claude Sonnet 4.5Small Model=Qwen3-VL-8B, Large Model=Claude Sonnet 4.5, Mode=Step-level Cascading2026.04 | 54.3 | 5.2 | 0.423 | 24.2 | 234 | 38.4 | 61.6 | |
| EvoCUA-8BModel Type=Standalone Small2026.04 | 43.3 | 2.6 | 0.022 | 30.9 | — | — | — | |
| Qwen3-VL-8BModel Type=Standalone Small2026.04 | 30.8 | 3.9 | 0.018 | 23.6 | — | — | — |