Benchmarks
Light
Benchmarks
MMBench (English)
Vision
unit % · normalized over [40, 96]
English split of MMBench.
#
Model
Source
Score
Normalized
1
Qwen3.5-397B-A17B
OPEN
self-reported
93.7%
2
Qwen3.5-122B-A10B
OPEN
self-reported
92.8%
3
Qwen3.5-Omni-Plus
CLOSED
self-reported
92.8%
4
Qwen3.6-35B-A3B
OPEN
self-reported
92.8%
5
Qwen3.6-27B
OPEN
self-reported
92.3%
6
ERNIE 4.5 VL 424B-A47B
OPEN
self-reported
92.0%
7
InternVL3-78B
OPEN
self-reported
89.0%
8
Qwen3.5-Omni-Flash
CLOSED
self-reported
88.8%
9
InternVL2.5-78B
OPEN
self-reported
88.3%
10
Qwen2.5-VL-72B
OPEN
self-reported
88.0%
11
Phi-4-multimodal (5.6B)
OPEN
self-reported
86.7%
12
Phi-4-multimodal-instruct
OPEN
self-reported
86.7%
13
Qwen2-VL-72B
OPEN
self-reported
86.5%
14
DeepSeek-VL2
OPEN
self-reported
83.1%
15
GLM-4V-9B
OPEN
self-reported
81.1%
16
MiniCPM-o 2.6
OPEN
self-reported
80.5%
17
MiniCPM-V 2.6
OPEN
self-reported
78.0%
18
MiniCPM-Llama3-V 2.5
OPEN
self-reported
77.2%
Score distribution
18 tracked results across the normalization window
40
96
Score vs. parameters
Open-weights models, log-x params
1B
10B
100B
1000B