Gemma 3 4B (IT): benchmark results
Google's instruction-tuned 4B Gemma 3 model (March 2025) taking image and text input, with 128K context and support for 140+ languages. Provider: Google. Released 2025-03-12. Access: Open.
Aggregate Rating 1451 ± 4, rank #992 of 1656 rated models, from 2214 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Bench2Drive-VL (BEV, No CoT) | 49.52 | Driving Score (0-100; Bench2Drive closed-loop protocol: rout | 100 |
| CVSBench - CVUSA Ground-to-Satellite | 71 | Accuracy (%) on CVUSA-subset ground-to-satellite questions: | 100 |
| EVALITA - MAIA-GEN | 58.39 | CPS | 100 |
| MEGA-Bench Task - Face Swap | 71.4 | Task Score (%) | 100 |
| MVX-Bench - Defeasible Entailment | 56 | Accuracy (%) on the 150 Defeasible Entailment questions (dec | 100 |
| TelcoAgent-Bench - Intent Recognition with Intent List (Arabic) | 86.4 | Intent recognition accuracy: mean cosine similarity x100 (-1 | 100 |
| TelcoAgent-Bench - Intent Recognition with Intent List (English) | 92.6 | Intent recognition accuracy: mean cosine similarity x100 (-1 | 100 |
| TelcoAgent-Bench - Resolution Accuracy (Arabic) | 84.7 | Resolution accuracy: mean cosine similarity x100 (-100 to 10 | 100 |
| LA Leaderboard - ClinDiagnosES | 65.33 | Semantic Answer Similarity (x100) | 98.5 |
| LA Leaderboard - TELEIA | 66.67 | Accuracy (%) | 98.5 |
| AGC-Bench - sdat | 0.67 | Dataset z-score | 97.5 |
| LA Leaderboard - BEC2016 Basque | 58.6 | F1 (%) | 97.1 |
Interactive version: theaggregate.ai/model?slug=gemma-3-4b-it · How It Works · Data refreshed daily, snapshot 2026-10-20.