List your company

Long Video Understanding on LongVideoBench (val)

80AccuracyLVAgent
61.866.52571.2575.975Aug 6, 2024Nov 28, 2024Mar 23, 2025Jul 16, 2025Nov 7, 2025Mar 2, 2026Jun 25, 2026

Evaluation results

MethodLinks
LVAgentAuxiliary Subtitles=true2025.1280————
VideoChat-A1Model=GPT-4o, Frame Budget (K)=384, Evaluation Protocol=Stratified random 25% subset2026.0377.2————
Gemini 2.5 ProAccess Type=API call only2026.0176.8————
Gemini-2.5-ProAccess=Proprietary2026.0276.8————
Gemini 2.5 ProCategory=API call only2026.0376.8————
Gemini 3 ProAccess Type=API call only2026.0175.9————
Gemini-3-ProAccess=Proprietary2026.0275.9————
Gemini 3 ProCategory=API call only2026.0375.9————
Seed-1.5-VL2025.1274.4————
AVPBackbone=Gemini-2.5-Pro2025.1273.4————
Gemini 2.5 FlashAccess Type=API call only2026.0173.1————
Gemini 2.5 FlashCategory=API call only2026.0373.1————
GPT-5Access Type=API call only2026.0172.6————
GPT-5Access=Proprietary2026.0272.6————
GPT-5Category=API call only2026.0372.6————
DeepVideoDiscovery (DVD)2025.1271.6————
LensWalkReasoner=o3, Observer=o32026.0370.6————
AVPBackbone=Gemini-2.5-Flash2025.1270.2————
HiMuModel=Gemini-2.5-Flash, Frame Budget (K)=16, Evaluation Protocol=Stratified random 25% subset2026.0370.13————
LensWalkReasoner=o3, Observer=GPT-4.12026.0369.9————
Gemini-2.5-Pro2025.1269.8————
GPT-5 miniAccess Type=API call only2026.0169.7————
GPT-5 miniCategory=API call only2026.0369.7————
LensWalkReasoner=GPT-5, Observer=GPT-52026.0368.8————
Deep Video Discovery2026.0368.6————
Molmo2-4BAccess Type=Molmo2 family2026.0168————
Molmo2-4BCategory=Fully open2026.0368————
VideoDetective (SeedVL-1.5)Param=20B(A), Frames=322026.0367.9————
LongVPO (512f) Stage 1Size=8B, Frames=512, Backbone=InternVideo2.52026.0267.8————
Molmo2-8BAccess Type=Molmo2 family2026.0167.5————
Molmo2-8BCategory=Fully open2026.0367.5————
LensWalkReasoner=o3, Observer=Qwen2.5-VL-72B2026.0367.3————
MolmoPoint-8BCategory=MolmoPoint2026.0367.3————
LongVPO (256f) Stage 2-iter1Size=8B, Frames=512, Backbone=InternVideo2.52026.0267.2————
LongVPO (256f) Stage 2-iter2Size=8B, Frames=512, Backbone=InternVideo2.52026.0267.2————
LongVPO (256f) Stage 1Size=8B, Frames=512, Backbone=InternVideo2.52026.0267————
AdaReTaKe2025.1267————
LongVPO (128f) Stage 1Size=8B, Frames=512, Backbone=InternVL32026.0266.8————
GPT-4o2024.0866.7————
GPT-4o2024.0866.7————
GPT-4ocategory=Proprietary2024.1066.7————
GPT-4o2024.1266.7————
GPT4-oSize=Undisclosed2026.0266.7————
GPT-4o#Frames=384, Model Type=Closed-source Models2026.0166.7————
GPT-4oSize=—2026.0166.7————
GPT-4o2025.1266.7————
GPT-4oParam=-, Frames=3842026.0366.7————
GPT-4o# Params=-2026.0666.7————
LongVPO (128f) Stage 2Size=8B, Frames=512, Backbone=InternVL2.52026.0266.6————
Eagle2.5-8BAccess Type=Open weights only2026.0166.4————
Eagle2.5-8BCategory=Open weights2026.0366.4————
Gemini-2.5-Flash2025.1266.2————
LongVPO (128f) Stage 2Size=8B, Frames=768, Backbone=InternVL32026.0266————
LongVPO (256f) Stage 1Size=8B, Frames=512, Backbone=InternVL2.52026.0266————
Keye-VL-1.5-8BAccess Type=Open weights only2026.0166————
Keye-VL-1.5-8BCategory=Open weights2026.0366————
GLM-4.1V-9BAccess Type=Open weights only2026.0165.7————
GLM-4.1V-9BCategory=Open weights2026.0365.7————
LongVPO (128f) Stage 1Size=8B, Frames=512, Backbone=InternVL2.52026.0265.4————
FrameOracleBackbone=Qwen3-VL-8B, Frames=128→27.82025.1065.2————
Claude Sonnet 4.5Access Type=API call only2026.0165.1————
Claude-Sonnet-4.5Access=Proprietary2026.0265.1————
HiMuModel=GPT-4o, Frame Budget (K)=16, Evaluation Protocol=Stratified random 25% subset2026.0365.1————
Claude Sonnet 4.5Category=API call only2026.0365.1————
VideoChat-FlashSize=7B, Frames=5122026.0264.7————
VideoChat-Flash-7BAccess Type=Open models2026.0164.7————
VideoChat-Flash-7BCategory=Fully open2026.0364.7————
GPT-4.12026.0364.6————
Vanilla MoEBackbone=KIMI-VL-A3B-INSTRUCT2026.0464.37————
MACS (Ours)Backbone=KIMI-VL-A3B-INSTRUCT2026.0464.24————
HiMuModel=Qwen3-VL-8B, Frame Budget (K)=16, Evaluation Protocol=Full benchmark2026.0364.19————
HiMuModel=InternVL-3.5-8B, Frame Budget (K)=16, Evaluation Protocol=Full benchmark2026.0364.11————
Gemini-1.5-Pro2024.0864————
Gemini-1.5-Procategory=Proprietary2024.1064————
Gemini-1.5-ProSize=Undisclosed2026.0264————
Gemini-1.5-pro#Frames=256, Model Type=Closed-source Models2026.0164————
Gemini-1.5-ProParam=-, Frames=2562026.0364————
Gemini-1.5-Pro# Params=-2026.0664————
FrameOracleBackbone=Qwen3-VL-8B, Frames=32→20.92025.1064————
MACS (w/o Expanded)Backbone=KIMI-VL-A3B-INSTRUCT2026.0463.95————
MiniCPM-V-4.5-8BAccess Type=Open weights only2026.0163.9————
LLaVA-VideoParam=72B, Frames=642026.0363.9————
MiniCPM-V-4.5-8BCategory=Open weights2026.0363.9————
SeedVL-1.5Param=20B(A), Frames=322026.0363.8————
Vanilla MoEBackbone=INTERNVL3.5-30B-A3B2026.0463.76————
Seed1.5VL-pro#Frames=32, Model Type=Closed-source Models2026.0163.7————
Molmo2-O-7BAccess Type=Molmo2 family2026.0163.7————
Molmo2-O-7BCategory=Fully open2026.0363.7————
MACS (Ours)Backbone=INTERNVL3.5-30B-A3B2026.0463.55————
VSLSModel=GPT-4o, Frame Budget (K)=32, Evaluation Protocol=Stratified random 25% subset2026.0363.4————
Qwen3-VL-8BFrames=322025.1063.3————
LLaVA-OneVision-72BModel Scale=72B2024.0863.2————
InternVideo2.5*Size=8B, Frames=5122026.0263.2————
MACS (w/o Expanded)Backbone=INTERNVL3.5-30B-A3B2026.0463.11————
ARIAactivated_params=3.9B (25.3B), category=Open-source2024.1063————
Qwen3-VL-4BAccess Type=Open weights only2026.0162.8————
Qwen3-VL-4BCategory=Open weights2026.0362.8————
InternVL2.5*Size=8B, Frames=5122026.0262.7————
AKSFrames=642025.1062.7————
BaselineRetained Tokens=All 64 x 169, Token Reduction Ratio=0%, Number of Frames=642026.0662.5——5965