Gemma 4 E2B: benchmark results

Google DeepMind's Gemma 4 E2B, a multimodal (text, image, audio) model in the five-size Gemma 4 family, with a 256K-token context, support for 140+ languages and configurable reasoning modes. Provider: Google. Released 2026-04-02. Access: Open.

Aggregate Rating 1434 ± 12, rank #4541 of 10318 rated models, from 548 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Jais 3 Model Cards - SimpleQA Verified (Arabic) - Non-hallucination Rate84.8Non-hallucination rate (%)100
Lara Mini Base Report (base-model long-context protocol) - HELMET 128k58.6Mean of four HELMET categories (recall, RAG, re-ranking, in-100
Lara Mini Base Report (base-model long-context protocol) - HELMET 32k66Mean of four HELMET categories (recall, RAG, re-ranking, in-100
Lara Mini Base Report (base-model long-context protocol) - ONERULER 128k84.7Mean of four ONERULER retrieval tasks in 26 languages (answe100
Lara Mini Base Report (base-model protocol) - Global PIQA (non-parallel), All Languages (0-shot)62.2Accuracy (%) averaged over all languages/variants of the ben96.3
Lara Mini Base Report (base-model protocol) - Global PIQA (parallel), All Languages (0-shot)28.3Accuracy (%) averaged over all languages/variants of the ben96.3
Lara Mini Base Report (base-model protocol) - T-HellaSwag, All Languages (0-shot)49.8Accuracy normalized by answer length (%) averaged over all l96.3
Lara Mini Base Report (base-model protocol) - XCOPA, All Languages (0-shot)64.1Accuracy (%) averaged over all languages/variants of the ben95.4
POLAR-Bench - Utility87.56Utility score (%): share of the task-relevant attributes the95.2
QIMMA - ArabLegalQA72.39BERTScore F1 (%)93.2
AA-Omniscience Hallucination Rate32.36Hallucination Rate (%)90.6
Lara Mini Base Report (base-model protocol) - XNLI, All Languages (0-shot)43.5Accuracy (%) averaged over all languages/variants of the ben89.8

Interactive version: theaggregate.ai/model?slug=gemma-4-e2b · How It Works · Data refreshed daily, snapshot 2026-11-14.