gemma-4-E4B-it: benchmark results
Google's Gemma 4 instruction-tuned on-device model (~4B effective params via per-layer embeddings) with text, image and audio input and a 128K context (April 2026). Provider: Google. Released 2026-04-02. Access: Open.
Aggregate Rating 1564 ± 6, rank #566 of 1656 rated models, from 980 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CBT-Audio (Audio + Transcript) | 0.72 | Mean absolute error on the 1-5 distress-intensity scale agai | 100 |
| CBT-Audio (Transcript Only) | 0.75 | Mean absolute error on the 1-5 distress-intensity scale agai | 100 |
| JMeetEval | 69 | Prompt-level strict accuracy (%) | 100 |
| JMeetEval - Instruction Level | 83.3 | Instruction-level strict accuracy (%) | 100 |
| JetBrains Mellum2.1 Card - WorkBench | 46.1 | Task completion (%) | 100 |
| Horangi 4 - BFCL | 70.95 | Accuracy (%) | 98.8 |
| EuroEval Italian Summarization - Ilpost SUM | 38.85 | chrF3++ (%) | 97.3 |
| Nejumi 4 - Toxicity - Violation Categories | 57.94 | Criteria met (%) | 97.2 |
| Nejumi 4 - BFCL - Irrelevance Detection | 95 | Accuracy (%) | 96.5 |
| EuroEval Portuguese NLU - SST-2 PT | 86.93 | Matthews correlation coefficient (%) | 95.9 |
| EuroEval Lithuanian NLU - Atsiliepimai | 49.2 | Matthews correlation coefficient (%) | 95 |
| SEA-HELM (Burmese) - Sentiment Analysis | 47.56 | Normalized Score | 94.7 |
Interactive version: theaggregate.ai/model?slug=gemma-4-e4b-it · How It Works · Data refreshed daily, snapshot 2026-10-20.