Gemma 3 4B: benchmark results
Google's open-weight Gemma 3 4B, a small multimodal model (March 2025). Provider: Google. Released 2025-03-12. Access: Open.
Aggregate Rating 1432 ± 6, rank #1089 of 1656 rated models, from 995 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Self-Explanation Simulatability (NSG) | 36.51 | Normalized Simulatability Gain (%) | 100 |
| TinyQABenchmark (core-en) | 86.5 | Exact Match | 100 |
| OpenVLM LLaVA-Bench - Detail | 96.6 | Score | 98.5 |
| OpenVLM MMT-Bench - Plant Recognition | 95 | Score (%) | 98.1 |
| Open LMM Reasoning - WeMath - InadequateGeneralization (Loose) | 15.6 | Accuracy (%) | 95.7 |
| OpenVLM MMT-Bench - Shape Recognition | 95 | Score (%) | 93.7 |
| OpenVLM MMT-Bench - Font Recognition | 40 | Score (%) | 93.2 |
| OpenVLM MMT-Bench - Tech Engineering | 50 | Score (%) | 93 |
| UrbanWell - Land Use Classification | 60 | Accuracy (%) of choosing the land-use category of a grid cel | 92.9 |
| OpenVLM MMT-Bench - Temporal Anticipation | 80 | Score (%) | 91.3 |
| OpenVLM MMBench V1.1 CN - Image Topic | 98.9 | Accuracy (%) | 90.3 |
| OpenVLM MMT-Bench - LVLM Response Judgement | 60 | Score (%) | 88.6 |
Interactive version: theaggregate.ai/model?slug=gemma-3-4b · How It Works · Data refreshed daily, snapshot 2026-10-20.