Google

Gemma 3 4B

Multimodal
Zero-eval
#3VQAv2 (val)
#3MMMU (val)

by Google

+
+
+
+
About

Gemma 3 4B was developed as a compact yet capable open-source model, designed to strike a balance between performance and resource efficiency. Built with 4 billion parameters and instruction tuning, it provides a practical option for applications requiring moderate capability with manageable computational costs.

+
+
+
+
Pricing Range
Input (per 1M)$0.02 -$0.02
Output (per 1M)$0.04 -$0.04
Providers1
+
+
+
+
Timeline
AnnouncedMar 12, 2025
ReleasedMar 12, 2025
Knowledge CutoffAug 1, 2024
+
+
+
+
Specifications
Training Tokens4.0T
Capabilities
Multimodal
+
+
+
+
License & Family
License
Gemma
Performance Overview
Performance metrics and category breakdown

Overall Performance

26 benchmarks
Average Score
53.0%
Best Score
90.2%
High Performers (80%+)
2

Performance Metrics

Max Context Window
262.1K
Avg Throughput
33.0 tok/s
Avg Latency
0ms
+
+
+
+
All Benchmark Results for Gemma 3 4B
Complete list of benchmark scores with detailed information
IFEval
text
0.90
90.2%
Self-reported
GSM8k
text
0.89
89.2%
Self-reported
DocVQA
multimodal
0.76
75.8%
Self-reported
MATH
text
0.76
75.6%
Self-reported
AI2D
multimodal
0.75
74.8%
Self-reported
BIG-Bench Hard
text
0.72
72.2%
Self-reported
HumanEval
text
0.71
71.3%
Self-reported
Natural2Code
text
0.70
70.3%
Self-reported
FACTS Grounding
text
0.70
70.1%
Self-reported
ChartQA
multimodal
0.69
68.8%
Self-reported
Showing 1 to 10 of 26 benchmarks
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+