Image Understanding on Multimodal Suite (TextVQA, ChartQA, MMStar, MMBench, MMVet, MME, RealWorldQA, COCO)
85.76TextVQA ScoreInternVL-3.5-30B-A3B-HF
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| InternVL-3.5-30B-A3B-HFk=8, mode=Default2025.11 | 85.76 | 84.08 | 62.49 | 83.81 | 69.93 | 2,312 | 64.77 | 69.3 | — | |
| Qwen3-VL-MoE-30B-A3B-Instructk=8, mode=Default2025.11 | 83.41 | 85.08 | 59.67 | 86.6 | 69.68 | 2,500 | 66.8 | 80.37 | — | |
| Qwen3-VL-MoE-30B-A3B-Instructk (Number of active experts)=8, Expert skipping ratio (rho)=02025.11 | 83.41 | 85.08 | 59.67 | 86.6 | 69.68 | 2,500 | 66.8 | 80.37 | 100 | |
| MC-MoEExpert skipping ratio (rho)=0.62025.11 | 82.51 | 79.37 | 56.48 | 86.12 | 69.37 | 2,438 | 62.01 | 76.82 | 97.09 | |
| DiEPExpert skipping ratio (rho)=0.62025.11 | 82.04 | 80.23 | 57.26 | 85.07 | 68.42 | 2,405 | 60.31 | 75.41 | 96.8 | |
| MoDESExpert skipping ratio (rho)=0.62025.11 | 81.82 | 82.48 | 58.61 | 86.17 | 69.95 | 2,493 | 63.92 | 76.55 | 99.22 | |
| MoDESExpert skipping ratio (rho)=0.732025.11 | 81.65 | 82.44 | 58.78 | 86.25 | 67.61 | 2,469 | 64.71 | 75.73 | 99.11 | |
| NAEEExpert skipping ratio (rho)=0.62025.11 | 81.2 | 79.41 | 55.39 | 84.18 | 68.61 | 2,348 | 59.67 | 78.09 | 95.61 | |
| MoDESBackbone=Qwen3-VL-MoE-30B-A3B-Instruct, Skip Experts=88%, rho=0.852025.11 | 80.97 | 78.84 | 58.18 | 85.57 | 67.75 | 2,403 | 64.58 | 74.66 | — | |
| MoDESExpert skipping ratio (rho)=0.852025.11 | 80.97 | 78.84 | 58.18 | 85.57 | 67.75 | 2,403 | 64.58 | 74.66 | 97.33 | |
| k = 3 (Default)k (Number of active experts)=3, Expert skipping ratio (rho)=0.62025.11 | 80.81 | 78.12 | 66.74 | 83.33 | 68.39 | 2,326 | 45.88 | 71.7 | 95.2 | |
| MoDESBackbone=InternVL-3.5-30B-A3B-HF, Skip Experts=88%, rho=0.852025.11 | 80.58 | 82 | 61.2 | 81.67 | 67.8 | 2,222 | 61.73 | 65.16 | — | |
| InternVL-3.5-GPT-OSS-20B-A4B-Preview-HFk=4, mode=Default2025.11 | 80.2 | 90.64 | 57.64 | 79.78 | 69.68 | 2,270 | 61.63 | 70.61 | — | |
| MC-MoEExpert skipping ratio (rho)=0.732025.11 | 80.13 | 78.41 | 57.02 | 85.32 | 67.22 | 2,286 | 61.83 | 74.49 | 95.11 | |
| DiEPExpert skipping ratio (rho)=0.732025.11 | 79.64 | 78.52 | 56.48 | 84.91 | 67.13 | 2,243 | 60.94 | 75.53 | 95.21 | |
| NAEEExpert skipping ratio (rho)=0.732025.11 | 78.42 | 77.28 | 54.64 | 81.34 | 65.58 | 2,208 | 61.75 | 77.31 | 93.25 | |
| MoDESBackbone=InternVL-3.5-GPT-OSS-20B-A4B-Preview-HF, Skip Experts=75%, rho=0.732025.11 | 77.93 | 89.6 | 56.48 | 78.14 | 66.33 | 2,206 | 60.64 | 68.32 | — | |
| k = 2 (Default)k (Number of active experts)=2, Expert skipping ratio (rho)=0.732025.11 | 77.54 | 69.6 | 62.38 | 80.5 | 61.33 | 2,060 | 55.56 | 82.77 | 92.03 | |
| DiEPBackbone=InternVL-3.5-GPT-OSS-20B-A4B-Preview-HF, Skip Experts=75%, rho=0.732025.11 | 77.31 | 86.24 | 48.18 | 74.26 | 58.07 | 2,109 | 60.25 | 62.08 | — | |
| MC-MoEBackbone=InternVL-3.5-GPT-OSS-20B-A4B-Preview-HF, Skip Experts=75%, rho=0.732025.11 | 76.49 | 84.53 | 46.25 | 73.68 | 56.83 | 2,137 | 61.07 | 60.42 | — | |
| MC-MoEBackbone=Qwen3-VL-MoE-30B-A3B-Instruct, Skip Experts=88%, rho=0.852025.11 | 74.87 | 71.43 | 50.74 | 75.42 | 61.35 | 2,168 | 60.41 | 68.15 | — | |
| MC-MoEExpert skipping ratio (rho)=0.852025.11 | 74.87 | 71.43 | 50.74 | 75.42 | 61.35 | 2,168 | 60.41 | 68.15 | 86.66 | |
| NAEEBackbone=InternVL-3.5-GPT-OSS-20B-A4B-Preview-HF, Skip Experts=75%, rho=0.732025.11 | 73.89 | 82.34 | 44.89 | 71.59 | 54.97 | 2,017 | 63.46 | 59.73 | — | |
| DiEPBackbone=Qwen3-VL-MoE-30B-A3B-Instruct, Skip Experts=88%, rho=0.852025.11 | 73.46 | 70.51 | 53.28 | 73.21 | 58.64 | 2,074 | 63.41 | 62.89 | — | |
| DiEPExpert skipping ratio (rho)=0.852025.11 | 73.46 | 70.51 | 53.28 | 73.21 | 58.64 | 2,074 | 63.41 | 62.89 | 85.3 | |
| NAEEBackbone=Qwen3-VL-MoE-30B-A3B-Instruct, Skip Experts=88%, rho=0.852025.11 | 72.41 | 65.83 | 48.88 | 73.62 | 54.52 | 1,984 | 58.62 | 60.37 | — | |
| NAEEExpert skipping ratio (rho)=0.852025.11 | 72.41 | 65.83 | 48.88 | 73.62 | 54.52 | 1,984 | 58.62 | 60.37 | 80.6 | |
| MC-MoEBackbone=InternVL-3.5-30B-A3B-HF, Skip Experts=88%, rho=0.852025.11 | 70.41 | 73.49 | 56.14 | 64.38 | 72.41 | 1,972 | 57.49 | 60.12 | — | |
| DiEPBackbone=InternVL-3.5-30B-A3B-HF, Skip Experts=88%, rho=0.852025.11 | 69.37 | 71.84 | 57.21 | 63.19 | 65.32 | 1,838 | 56.38 | 55.78 | — | |
| InternVL-3.5-GPT-OSS-20B-A4B-Preview-HFk=1, Skip Experts=75%, rho=0.732025.11 | 68.74 | 79.72 | 45.77 | 67.63 | 48.49 | 1,833 | 53.2 | 60.7 | — | |
| NAEEBackbone=InternVL-3.5-30B-A3B-HF, Skip Experts=88%, rho=0.852025.11 | 66.24 | 68.32 | 50.14 | 64.37 | 49.52 | 1,802 | 55.23 | 50.64 | — | |
| Qwen3-VL-MoE-30B-A3B-Instructk=1, Skip Experts=88%, rho=0.852025.11 | 60.71 | 52.16 | 31.63 | 54.9 | 28.07 | 1,590 | 52.42 | 45.64 | — | |
| k = 1 (Default)k (Number of active experts)=1, Expert skipping ratio (rho)=0.852025.11 | 60.71 | 52.16 | 31.63 | 54.9 | 28.07 | 1,590 | 52.42 | 45.64 | 60.11 | |
| InternVL-3.5-30B-A3B-HFk=1, Skip Experts=88%, rho=0.852025.11 | 58.49 | 46.24 | 42.27 | 51.74 | 35.05 | 1,683 | 51.44 | 26.01 | — |