Benchmarks
Light
Benchmarks
VQAv2
Vision
unit % · normalized over [40, 87]
Open-ended visual question answering.
#
Model
Source
Score
Normalized
1
Reka Edge (2026)
OPEN
self-reported
88.4%
2
Molmo 2 8B
OPEN
self-reported
87.0%
3
Molmo 72B
OPEN
self-reported
86.5%
4
CogVLM-17B
OPEN
self-reported
82.3%
5
Idefics2 8B
OPEN
self-reported
81.2%
6
Pixtral Large (124B)
OPEN
self-reported
80.9%
7
Gemini 1.5 Pro
CLOSED
self-reported
80.2%
8
Gemini 1.5 Flash
CLOSED
self-reported
80.1%
9
Qwen-VL
OPEN
self-reported
78.8%
10
Pixtral 12B
OPEN
self-reported
78.6%
11
Reka Flash (21B)
CLOSED
self-reported
78.4%
12
Llama 3.2 90B Vision
OPEN
self-reported
78.1%
13
Reka Core
CLOSED
self-reported
78.1%
14
Gemini 1.0 Ultra
CLOSED
self-reported
77.8%
15
Llama 3.2 11B Vision
OPEN
self-reported
75.2%
16
Gemma 3 12B
OPEN
self-reported
71.6%
17
Gemini 1.0 Pro
CLOSED
self-reported
71.2%
18
Gemini 1.0 Nano
CLOSED
self-reported
67.5%
19
Gemma 3 4B
OPEN
self-reported
62.4%
20
IDEFICS 80B
OPEN
self-reported
60.0%
Score distribution
20 tracked results across the normalization window
40
87
Score vs. parameters
Open-weights models, log-x params
1B
10B
100B
1000B