Video Understanding on Video-MME (Accuracy w/o subtitles, Accuracy w/ subtitles)
65.6Accuracy (w/o subtitles)InternVL3-8B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| InternVL3-8BFrames=32, Tokens/Frame=256, Vision Params=304M, Vision Latency (ms)=144.5, LLM Latency (ms)=669.8, Total Latency (ms)=814.52026.05 | 65.6 | 67.4 | |
| InternVL3-8B + LiteFrameFrames=256 (8×), Tokens/Frame=16, Vision Params=87M, Vision Latency (ms)=204.0, LLM Latency (ms)=327.4, Total Latency (ms)=532.3 (-34.6%)2026.05 | 65.1 | 68.5 | |
| InternVL3-8B + LiteFrameFrames=128 (8×), Tokens/Frame=16, Vision Params=87M, Vision Latency (ms)=105.3, LLM Latency (ms)=166.6, Total Latency (ms)=272.6 (-32.5%)2026.05 | 63.9 | 66.8 | |
| InternVL3-8BFrames=16, Tokens/Frame=256, Vision Params=304M, Vision Latency (ms)=74.0, LLM Latency (ms)=329.3, Total Latency (ms)=403.62026.05 | 61.9 | 64 | |
| InternVL3-8B + LiteFrameFrames=64 (8×), Tokens/Frame=16, Vision Params=87M, Vision Latency (ms)=54.8, LLM Latency (ms)=94.9, Total Latency (ms)=150.1 (-28.0%)2026.05 | 61 | 64.2 | |
| InternVL3-8B + FastVIDFrames=128 (4×), Tokens/Frame=16, Vision Params=304M, Vision Latency (ms)=625.8, LLM Latency (ms)=168.9, Total Latency (ms)=794.9 (-2.4%)2026.05 | 60.4 | 66.1 | |
| InternVL3-8B + FastVIDFrames=64 (4×), Tokens/Frame=16, Vision Params=304M, Vision Latency (ms)=310.6, LLM Latency (ms)=95.4, Total Latency (ms)=406.2 (+0.6%)2026.05 | 59.3 | 63.4 | |
| InternVL3-8BFrames=8, Tokens/Frame=256, Vision Params=304M, Vision Latency (ms)=40.0, LLM Latency (ms)=167.3, Total Latency (ms)=208.42026.05 | 59 | 60 | |
| InternVL3-8B + FastVIDFrames=32 (4×), Tokens/Frame=16, Vision Params=304M, Vision Latency (ms)=161.7, LLM Latency (ms)=63.0, Total Latency (ms)=224.8 (+7.9%)2026.05 | 58.7 | 62.5 |