Gemma 3n E2B: benchmark results
Google's Gemma 3n E2B, a compact checkpoint derived from Gemma 3n E4B in the Gemma 3n family, released under the Gemma license; the linked card provides no further training details here. Provider: Google. Released 2025-06-26. Access: Open.
Aggregate Rating 1412 ± 19, rank #1158 of 1635 rated models, from 73 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| PARSA-Bench - Music Understanding | 37 | Accuracy (0-1 times 100) on 496 four-option questions about | 71.4 |
| PARSA-Bench - Speech Translation (Persian to English) | 64 | COMET (0-1 times 100) of English translations of 500 spoken | 71.4 |
| AudioProcessBench - Acoustic Attribute Errors | 45.2 | Type-conditioned PRMScore (0-100): detection of steps with a | 60 |
| AudioProcessBench - Existence Errors | 53.6 | Type-conditioned PRMScore (0-100): detection of steps with e | 60 |
| AudioProcessBench - All-Error Identification | 49.2 | PRMScore (0-100) of the binary correctness judgment on every | 50 |
| AudioProcessBench - Cross-Modal Binding Errors | 68.1 | Type-conditioned PRMScore (0-100): detection of steps with c | 50 |
| AudioProcessBench - Temporal Errors | 46.3 | Type-conditioned PRMScore (0-100): detection of steps with t | 50 |
| AudioProcessBench | 45.6 | Overall score (0-100): mean of the step-correctness, error-t | 40 |
| AudioProcessBench (Zero-shot) | 44.2 | Overall score (0-100): mean of the step-correctness, error-t | 40 |
| AudioProcessBench - Reasoning Errors | 47.7 | Type-conditioned PRMScore (0-100): detection of steps with r | 40 |
| AudioProcessBench - Semantic Errors | 47.6 | Type-conditioned PRMScore (0-100): detection of steps with s | 40 |
| PRISM-VLM - Instruction Following | 69.7 | Per-axis mean score (%; mean compliance with 1-3 verifiable | 36.6 |
Interactive version: theaggregate.ai/model?slug=gemma-3n-e2b · How It Works · Data refreshed daily, snapshot 2026-11-05.