Video Question Answering on VideoMME (Short, Median, Long, Avg. Metrics)
76Short VQA AccuracyFull Tokens
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Full TokensRetained Ratio (%)=100, Model Backbone=Qwen2.5-Omni-7B2026.07 | 76 | 68.33 | 64.89 | 69.74 | |
| Gemini 3.0 ProAccess=Proprietary, Protocol=Prism2026.07 | 75.11 | 69.67 | 60.78 | 68.52 | |
| OmniFocusRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-7B2026.07 | 74.78 | 68 | 65.44 | 69.41 | |
| Seed 2.0 ProAccess=Proprietary, Protocol=Prism2026.07 | 74.67 | 68 | 57.22 | 66.63 | |
| Full TokensRetained Ratio (%)=100, Model Backbone=Qwen2.5-Omni-3B2026.07 | 74.11 | 65.56 | 60.11 | 66.59 | |
| OmniZipRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-7B2026.07 | 73.67 | 66 | 64.33 | 68 | |
| CuReAccess=Open-source, Protocol=Prism2026.07 | 73.67 | 66.33 | 56.33 | 65.44 | |
| OmniFocusRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-7B2026.07 | 73.22 | 66.89 | 64.56 | 68.22 | |
| Qwen3.6-35B-A3BAccess=Open-source, Protocol=Prism2026.07 | 73.11 | 65.67 | 44.67 | 61.15 | |
| DyCokeRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-7B2026.07 | 73.01 | 64.33 | 65.33 | 67.99 | |
| V-LynX-7BModel Category=Task-specific models ≥ 7B, ΔParams.=195.0M2026.05 | 73 | 61.2 | 53.8 | 62.7 | |
| OmniZipRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-7B2026.07 | 72.89 | 66.33 | 64.67 | 67.96 | |
| RandomRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-7B2026.07 | 72.11 | 65 | 65.67 | 67.59 | |
| DyCokeRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-7B2026.07 | 72.01 | 63.33 | 64.33 | 66.99 | |
| OmniFocusRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-3B2026.07 | 71.89 | 62.56 | 60 | 64.81 | |
| RandomRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-7B2026.07 | 71.44 | 65.22 | 65.33 | 67.33 | |
| OmniFocusRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-3B2026.07 | 71.22 | 62.89 | 61 | 65.04 | |
| PAVE-7BModel Category=Task-specific models ≥ 7B, ΔParams.=500.5M2026.05 | 71.1 | 59.4 | 49.2 | 59.9 | |
| OmniZipRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-3B2026.07 | 70.56 | 63.67 | 59.89 | 64.7 | |
| OmniZipRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-3B2026.07 | 70.56 | 63.22 | 59.78 | 64.52 | |
| LLaVA-OV-7BModel Category=Task-specific models ≥ 7B2026.05 | 70.1 | 56.6 | 48.9 | 58.2 | |
| DyCokeRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-3B2026.07 | 70.01 | 62.1 | 59.78 | 64 | |
| RandomRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-3B2026.07 | 67.89 | 62.56 | 59.56 | 63.33 | |
| DyCokeRetained Ratio (%)=35, Model Backbone=Qwen2.5-Omni-3B2026.07 | 67.89 | 62.56 | 59.56 | 63.33 | |
| RandomRetained Ratio (%)=25, Model Backbone=Qwen2.5-Omni-3B2026.07 | 67.67 | 63.22 | 59.78 | 63.56 | |
| V-LynX-0.5BModel Category=Task-specific models < 7B, ΔParams.=68.7M2026.05 | 63.1 | 50.7 | 44.6 | 52.8 | |
| Qwen3-VL-235B-A22B-InstructAccess=Open-source, Protocol=Prism2026.07 | 61.33 | 53.56 | 46.56 | 53.81 | |
| Qwen3-VL-30B-A3B-InstructAccess=Open-source, Protocol=Prism2026.07 | 61 | 51.33 | 43.78 | 52.04 | |
| PAVE-0.5BModel Category=Task-specific models < 7B, ΔParams.=371.4M2026.05 | 57.8 | 42.7 | 37.4 | 46 | |
| OwlCap-7BAccess=Open-source, Protocol=Prism2026.07 | 53.78 | 41.56 | 38.33 | 44.56 | |
| LLaVA-OV-0.5BModel Category=Task-specific models < 7B2026.05 | 53.4 | 41.2 | 37.3 | 44 | |
| Tarsier2-7BAccess=Open-source, Protocol=Prism2026.07 | 40.44 | 30.78 | 28.33 | 33.19 |