Streaming Video Understanding Benchmarks
MOSS-VL-Realtime vs. open-source streaming baselines, subset by subset
OVOBench
MOSS-VL-Realtime
AURA
Streamo-7B
ViSpeak-7B
40
50
60
70
80
90
62.1
55.8
57.0
54.2
FAR
Forward Active Responding
72.6
60.4
49.2
57.52
BT
Backward Tracing
75.9
79.8
67.4
66.28
RTVP
Real-Time Visual Perception
70.2
65.3
57.9
59.3
Avg
OmniMMI
MOSS-VL-Realtime
AURA
0
10
20
30
40
50
60
70
66.0
37.5
PA
Proactive Alerting
21.67
24.0
SG
Dynamic State Grounding
10.67
7.7
MD
Multi-turn Dependency
33.5
32.0
AP
Action Prediction
31.5
26.0
SI
Speaker Identification
32.7
25.4
Avg
StreamingBench
MOSS-VL-Realtime
AURA
ViSpeak-7B
0
10
20
30
40
50
60
70
80
90
60.0
53.2
50.8
PO
Proactive Output
82.9
83.2
70.44
RT
Real-Time Visual Understanding
56.4
59
43.9
CTX
Contextual Understanding
50.4
57.2
39.2
SQA
Sequential Question Answering
62.4
63.2
51.1
Avg
ProactiveVideoQA
MOSS-VL-Realtime
M
MMDuet2-RL
M
MMDuet+rm.ass.turns
M
MMDuet
0
10
20
30
40
50
60
55.41
53.3
M
2-RL
43.5
M
+rm
38.9
M
base
WEB
Web Video QA
47.76
33.6
M
2-RL
52.2
M
+rm
46.0
M
base
EGO
Egocentric Video QA
50.2
43.4
M
2-RL
32.6
M
+rm
21.1
M
base
TV
TV Series QA
0
10
20
30
40
50
60
35.33
28.9
M
2-RL
42.5
M
+rm
27.4
M
base
VAD
Video Anomaly Detection
47.2
39.8
M
2-RL
42.7
M
+rm
33.4
M
base
Avg