Gemma 4 26B A4B: benchmark results

Google's Apache-2.0 Gemma 4 MoE (April 2026) with 26B total and ~4B active parameters, multimodal with context up to 256K. Provider: Google. Released 2026-04-02. Access: Open.

Unified ELO 1571 ± 1, rank #357 of 1392 rated models, from 354 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
Kaggle FACTS Grounding80.87Score (%)100
EuroEval Bulgarian NLU - Cinexio57.15Sentiment classification Score (%)92.4
EuroEval Icelandic NLU - NQII57.2Reading comprehension Score (%)90.6
MathVision82.4Overall Accuracy (%)88.8
Chatbot Arena (Text - Math)1468Arena Score88.5
EuroEval Estonian NLU - MultiWikiQA ET70.19Reading comprehension Score (%)87.1
EuroEval Czech NLU - SQAD68.08Reading comprehension Score (%)85.5
EuroEval Ukrainian NLU - MultiWikiQA UK62.69Reading comprehension Score (%)85.4
EuroEval German NLU - Germanquad62.82Reading comprehension Score (%)83.8
Chatbot Arena (Text - Instruction Following)1439Arena Score83.1
EuroEval Polish NLU - Polemo291.51Sentiment classification Score (%)82.8
EuroEval Serbian NLU - MultiWikiQA SR62.56Reading comprehension Score (%)81.9

Interactive version: theaggregate.ai/model?slug=gemma-4-26b-a4b · How It Works · Data refreshed daily, snapshot 2026-09-05.