Task Accuracy on OSWorld
41.49Task AccuracyKIVI
Evaluation Results
| Method | Links | |
|---|---|---|
| KIVIModel=Qwen3-VL-32B-Thinking2026.05 | 41.49 | |
| TriAxialKV INT4Model=Qwen3-VL-32B-Thinking2026.05 | 41.43 | |
| TriAxialKV MixedModel=Qwen3-VL-32B-Thinking2026.05 | 40.59 | |
| SGLang FP4Model=Qwen3-VL-32B-Thinking2026.05 | 40.54 | |
| SGLang BF16Model=Qwen3-VL-32B-Thinking2026.05 | 39.2 | |
| TriAxialKV MixedModel=Qwen3-VL-8B-Thinking2026.05 | 35.78 | |
| KIVIModel=Qwen3-VL-8B-Thinking2026.05 | 34.61 | |
| SGLang BF16Model=Qwen3-VL-8B-Thinking2026.05 | 34.02 | |
| TriAxialKV INT4Model=Qwen3-VL-8B-Thinking2026.05 | 33.92 | |
| SGLang FP4Model=Qwen3-VL-8B-Thinking2026.05 | 29.52 | |
| TriAxialKV INT4Model=InternVL3.5-38B2026.05 | 13.11 | |
| TriAxialKV MixedModel=InternVL3.5-38B2026.05 | 12.77 | |
| KIVIModel=InternVL3.5-38B2026.05 | 12.26 | |
| SGLang BF16Model=InternVL3.5-38B2026.05 | 12.22 | |
| SGLang FP4Model=InternVL3.5-38B2026.05 | 12.13 |