Gemma 3 12B (IT): benchmark results

Instruction-tuned Gemma 3 12B checkpoint. Provider: Google. Released 2025-03-12. Access: Open.

Aggregate Rating 1534 ± 4, rank #664 of 1656 rated models, from 2431 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Alexandria - English to Lebanese Arabic (Single Turn, Gender Only)21.3spBLEU (0-100; SacreBLEU with the FLORES-200 SentencePiece t100
Alexandria - English to Lebanese Arabic (Single Turn, No Metadata)20.82spBLEU (0-100; SacreBLEU with the FLORES-200 SentencePiece t100
BBC News Video Captioning - Entity Fidelity31.4Entity Fidelity Score (0-1, times 100): per-clip harmonic me100
BBC News Video Captioning - Text Similarity43.87Text similarity (cosine, -1 to 1, times 100) between all-Min100
BehaviorBench (Behavioral Traces) - Belief Confidence Error (DirectGen)0.11Confidence mean absolute error on a 0 to 1 scale: gap betwee100
BehaviorBench (Behavioral Traces) - Trade Amount Error (DirectGen)8.34Median absolute error of the predicted trade amount (paper u100
BehaviorBench (Behavioral Traces) - Trade Amount Error (RetrievalGen)8.29Median absolute error of the predicted trade amount (paper u100
Belebele European - Albanian87.22Accuracy (%, 0-shot)100
Belebele European - Estonian86Accuracy (%, 0-shot)100
Belebele European - Greek89.11Accuracy (%, 0-shot)100
Belebele European - Icelandic83.44Accuracy (%, 0-shot)100
Belebele European - Lithuanian86.67Accuracy (%, 0-shot)100

Interactive version: theaggregate.ai/model?slug=gemma-3-12b-it · How It Works · Data refreshed daily, snapshot 2026-10-20.