Streaming Video Understanding Benchmarks

MOSS-VL-Realtime vs. open-source streaming baselines, subset by subset

OVOBench

MOSS-VL-RealtimeMOSS-VL-Realtime
AURAAURA
Streamo-7BStreamo-7B
ViSpeak-7BViSpeak-7B
40 50 60 70 80 90
62.1MOSS-VL-Realtime
55.8AURA
57.0Streamo-7B
54.2ViSpeak-7B
FARForward Active Responding
72.6MOSS-VL-Realtime
60.4AURA
49.2Streamo-7B
57.52ViSpeak-7B
BTBackward Tracing
75.9MOSS-VL-Realtime
79.8AURA
67.4Streamo-7B
66.28ViSpeak-7B
RTVPReal-Time Visual Perception
70.2MOSS-VL-Realtime
65.3AURA
57.9Streamo-7B
59.3ViSpeak-7B
Avg

OmniMMI

MOSS-VL-RealtimeMOSS-VL-Realtime
AURAAURA
0 10 20 30 40 50 60 70
66.0MOSS-VL-Realtime
37.5AURA
PAProactive Alerting
21.67MOSS-VL-Realtime
24.0AURA
SGDynamic State Grounding
10.67MOSS-VL-Realtime
7.7AURA
MDMulti-turn Dependency
33.5MOSS-VL-Realtime
32.0AURA
APAction Prediction
31.5MOSS-VL-Realtime
26.0AURA
SISpeaker Identification
32.7MOSS-VL-Realtime
25.4AURA
Avg

StreamingBench

MOSS-VL-RealtimeMOSS-VL-Realtime
AURAAURA
ViSpeak-7BViSpeak-7B
0 10 20 30 40 50 60 70 80 90
60.0MOSS-VL-Realtime
53.2AURA
50.8ViSpeak-7B
POProactive Output
82.9MOSS-VL-Realtime
83.2AURA
70.44ViSpeak-7B
RTReal-Time Visual Understanding
56.4MOSS-VL-Realtime
59AURA
43.9ViSpeak-7B
CTXContextual Understanding
50.4MOSS-VL-Realtime
57.2AURA
39.2ViSpeak-7B
SQASequential Question Answering
62.4MOSS-VL-Realtime
63.2AURA
51.1ViSpeak-7B
Avg

ProactiveVideoQA

MOSS-VL-RealtimeMOSS-VL-Realtime
MMMDuet2-RL
MMMDuet+rm.ass.turns
MMMDuet
0 10 20 30 40 50 60
55.41MOSS-VL-Realtime
53.3M2-RL
43.5M+rm
38.9Mbase
WEBWeb Video QA
47.76MOSS-VL-Realtime
33.6M2-RL
52.2M+rm
46.0Mbase
EGOEgocentric Video QA
50.2MOSS-VL-Realtime
43.4M2-RL
32.6M+rm
21.1Mbase
TVTV Series QA
0 10 20 30 40 50 60
35.33MOSS-VL-Realtime
28.9M2-RL
42.5M+rm
27.4Mbase
VADVideo Anomaly Detection
47.2MOSS-VL-Realtime
39.8M2-RL
42.7M+rm
33.4Mbase
Avg