Gemma 3n E4B: benchmark results
Google's effective-4B on-device Gemma 3n model (June 2025) taking image, audio, video, and text within roughly 3GB of memory. Provider: Google. Released 2025-06-26. Access: Open.
Aggregate Rating 1472 ± 15, rank #878 of 1656 rated models, from 193 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EuroEval Finnish NLU - Scandisent FI | 93.19 | Matthews correlation coefficient (%) | 97.6 |
| EuroEval German NLU - Germanquad | 68.37 | Character F1 (%) | 95.9 |
| EuroEval Finnish NLU - Tydiqa FI | 75.84 | Character F1 (%) | 94.6 |
| EuroEval Icelandic NLU - NQII | 60.23 | Character F1 (%) | 94.5 |
| EuroEval Portuguese NLU - MultiWikiQA PT | 77.48 | Character F1 (%) | 94.5 |
| EuroEval French NLU - FQuAD | 74.51 | Character F1 (%) | 93.6 |
| EuroEval Norwegian NLU - Norquad | 74.34 | Character F1 (%) | 93.5 |
| EuroEval Swedish NLU - Swerec | 79.51 | Matthews correlation coefficient (%) | 93.4 |
| EuroEval Spanish NLU - MLQA ES | 67.25 | Character F1 (%) | 89.7 |
| EuroEval Faroese NLU - FoQA | 74.21 | Character F1 (%) | 86 |
| PARSA-Bench - Speech Translation (Persian to English) | 68 | COMET (0-1 times 100) of English translations of 500 spoken | 85.7 |
| EuroEval Dutch NLU - SQuAD NL | 77.8 | Character F1 (%) | 85.5 |
Interactive version: theaggregate.ai/model?slug=gemma-3n-e4b · How It Works · Data refreshed daily, snapshot 2026-10-20.