Gemma 4 E4B: benchmark results
Google's open-weight Gemma 4 E4B (April 2026), a compact multimodal member of the Gemma 4 family handling text, image and audio input, built on Gemini 3 research under Apache 2.0. Provider: Google. Released 2026-04-02. Access: Open.
Aggregate Rating 1505 ± 9, rank #767 of 1656 rated models, from 536 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CanLegalRAGBench (Oracle Retrieval) - Groundedness | 79.83 | Groundedness (%; share of the answer's atomic claims support | 100 |
| PRISM-VLM - Hallucination Resistance | 74.6 | Per-axis mean score (%; share of trap questions (absent enti | 97.6 |
| AA-Omniscience Hallucination Rate | 30.94 | Hallucination Rate (%) | 91.5 |
| POLAR-Bench - Utility | 86.29 | Utility score (%): share of the task-relevant attributes the | 90.5 |
| QIMMA - ArabLegalQA | 72.07 | BERTScore F1 (%) | 90.4 |
| aiXamine - Safety Alignment | 98.1 | Service score (0-100; normalized pass rate) | 87.8 |
| POLAR-Bench | 90.87 | Overall Mean (self-reported) | 85.7 |
| P3B3 | 94.4 | LLM pt-PT (↑) (self-reported) | 84.2 |
| aiXamine - Fairness & Bias | 65.8 | Service score (0-100; normalized pass rate) | 75.6 |
| MathVision | 59.5 | Overall Accuracy (%) | 75.2 |
| SEA-HELM | 66.3 | Mean Score (%) | 73.7 |
| MulRobBench | 46.29 | Semantic protocol-decision score (%; aggregate over constrai | 73.3 |
Interactive version: theaggregate.ai/model?slug=gemma-4-e4b · How It Works · Data refreshed daily, snapshot 2026-10-20.