List your company
ResearchFormerly WizWand

Video Understanding on MLVU (Accuracy)

81.5AccuracyRETOOL-VIDEO
59.76465.40771.0576.693Aug 3, 2025Sep 26, 2025Nov 19, 2025Jan 12, 2026Mar 7, 2026Apr 30, 2026Jun 23, 2026
Updated 4mo ago

Evaluation results

MethodLinks
RETOOL-VIDEOSize=9B2026.0581.5
Gemini-2.5-Flash-Lite2026.0578.5
AdaQEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0676.4
AdaQEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0675
Q-FrameEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0674.7
Gemini 1.5 ProSetting=best-setting numbers from official reports2026.0574.3
AdaQEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0674.2
AKSEmbedding Model=BLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0674.2
Top-kEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0673.2
AdaQEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0673.1
Video-LLaMA3-7BEvaluation Protocol=vanilla offline2026.0573
InternVL3.5-30B-A3BSize=30B2026.0573
Claude-Sonnet-4.52026.0572.8
Q-FrameEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0672.8
AdaQEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0672.6
Q-FrameEmbedding Model=LongCLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0672.4
AKSEmbedding Model=BLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0671.8
AdaQEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0671.8
OneClip-RAGEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen3-VL-8B2026.0671.5
LLaVA-Video-7BBase Model=LLaVA-Video-7B, Input Frames=642026.0471.4
Top-kEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0671.2
OneClip-RAGEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0671.2
Qwen3-VL-8BFrames=64, MLLM Backbone=Qwen3-VL-8B2026.0671
LLaVA-Video-7BEvaluation Protocol=vanilla offline2026.0570.8
InternVL3-8B + LiteFrameFrames=256 (8×)2026.0570.7
Top-kEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0670.7
Tango†Base Model=LLaVA-Video-7B, Retention Ratio=20%, Intra-LLM=true, Input Frames=642026.0470.4
LLaVA-Video 7BRetention Ratio R=100%, # Newline Tokens M=8322026.0570.3
BOLTEmbedding Model=CLIP, Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0670.3
E-VRAGEmbedding Model=VLM (2B), Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0670.2
NVILASize=8B2026.0570.1
InternVL3-8BFrames=322026.0569.6
AKSEmbedding Model=BLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0669.6
LLaVA-Video-7BFrames=64, MLLM Backbone=LLaVA-Video-7B2026.0669.5
AdaQEmbedding Model=LongCLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0669.4
InternVL3-8B + FastVIDFrames=128 (4×)2026.0569.3
GPT-5-Nano-2025-08-072026.0569.2
FRAGEmbedding Model=MLLM (7B), Frames=64, MLLM Backbone=LLaVA-Video-7B2026.0669.2
InternVL-V2.5-8BEvaluation Protocol=vanilla offline2026.0568.9
TangoBase Model=LLaVA-Video-7B, Retention Ratio=20%, Input Frames=642026.0468.7
OneClip-RAGEmbedding Model=CLIP, Frames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0668.6
Q-FrameEmbedding Model=LongCLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0668.5
AdaQEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0668.5
HoliTomBase Model=LLaVA-Video-7B, Retention Ratio=20%, Input Frames=642026.0468.2
FastVIDBase Model=LLaVA-Video-7B, Retention Ratio=20%, Input Frames=642026.0468.1
VisionZipBase Model=LLaVA-Video-7B, Retention Ratio=20%, Input Frames=642026.0467.7
Tango†Base Model=LLaVA-Video-7B, Retention Ratio=10%, Intra-LLM=true, Input Frames=642026.0467.6
OneClip-RAGEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0667.6
StreamAgent-7B#Frames=1fps2025.0867.2
Qwen2.5-VL-7B#Frames=1fps2025.0866.9
Top-kEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0666.9
AKSEmbedding Model=BLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0666.8
BOLTEmbedding Model=CLIP, Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0666.8
TangoBase Model=LLaVA-Video-7B, Retention Ratio=10%, Input Frames=642026.0466.7
InternVL3-8B + LiteFrameFrames=128 (8×)2026.0566.7
InternVL3-8BFrames=162026.0566.4
Flash-VStreamSize=7B2026.0566.3
HoliTomBase Model=LLaVA-Video-7B, Retention Ratio=10%, Input Frames=642026.0466.2
InternVL-V3.5-8BEvaluation Protocol=vanilla offline2026.0566.2
EvoStreaming-8B-InternVL-V3.5Evaluation Protocol=vanilla offline2026.0566
InternVL3-8B + LiteFrameFrames=64 (8×)2026.0565.7
DyCokeRetention Ratio R=32.1%, # Newline Tokens M=2562026.0565.7
Qwen2.5-VL-7BFrames=64, MLLM Backbone=Qwen2.5-VL-7B2026.0665.5
TimeChat-Online-7B#Frames=1fps2025.0865.4
LongVU-7BEvaluation Protocol=vanilla offline2026.0565.4
FRAGEmbedding Model=MLLM (7B), Frames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0665.3
LLaVA-OV 7BRetention Ratio R=100%2026.0565.2
Video-Thinker-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0565.2
FastVIDBase Model=LLaVA-Video-7B, Retention Ratio=10%, Input Frames=642026.0465.1
InternVL3-8B + FastVIDFrames=64 (4×)2026.0565.1
ParaVT-8BSetting=tool-augmented setting (<think>→<tool_call>→<answer>)2026.0565
LLaVA-OneVision-7BEvaluation Protocol=vanilla offline2026.0564.7
GPT-4oEvaluation Protocol=vanilla offline2026.0564.6
GPT-4o2026.0564.6
DynaTokRetention Ratio R=24.9%, # Newline Tokens M=8322026.0564.6
GPT-4oSetting=best-setting numbers from official reports2026.0564.6
Qwen2.5-VL-7BBase Model=Qwen2.5-VL-7B, Input Frames=322026.0464.4
VideoMindSize=7B2026.0564.4
InternVL-V2-8BEvaluation Protocol=vanilla offline2026.0564
InternVL3-8B + FastVIDFrames=32 (4×)2026.0564
FastVRetention Ratio R=100%/25%, # Newline Tokens M=832/568.72026.0563.9
LLaVA-OneVision-7BFrames=32, MLLM Backbone=LLaVA-OneVision-7B2026.0663.7
DynaTokRetention Ratio R=9.5%, # Newline Tokens M=8322026.0563.2
VisionZipBase Model=LLaVA-Video-7B, Retention Ratio=10%, Input Frames=642026.0462.9
TimeChat-Online-7BEvaluation Protocol=vanilla offline2026.0562.9
DynaTokRetention Ratio R=25%2026.0562.5
InternVL3-8BFrames=82026.0562.2
DyCokeRetention Ratio R=32.5%2026.0562.1
DynaTokRetention Ratio R=19.9%2026.0562.1
DynaTokRetention Ratio R=14.8%2026.0562.1
VisionZipRetention Ratio R=24.9%, # Newline Tokens M=642026.0562.1
Dispider-7B#Frames=1fps2025.0861.7
Dispider-7BEvaluation Protocol=vanilla offline2026.0561.7
Video-R1-7BSetting=reasoning-enhanced setting (<think>→<answer>)2026.0561.6
FastVRetention Ratio R=100%/25%2026.0561.5
FastVRetention Ratio R=100%/20%2026.0561.2
TangoBase Model=Qwen2.5-VL-7B, Retention Ratio=20%, Input Frames=322026.0461
Kangaroo-7BEvaluation Protocol=vanilla offline2026.0561
KangarooSize=8B2026.0561
VisionZipBase Model=Qwen2.5-VL-7B, Retention Ratio=20%, Input Frames=322026.0460.6