Benchmark Registry

BIG-Bench Extra Hard

BIG-Bench Extra Hard results reported for 5 models across 1 version. Updated 2026-09-24.

Report an issue

Versions

1 version · Metric: Micro-average accuracy

VersionReleasedModelsResults
BIG-Bench Extra Hard Original — micro averageLatestFebruary 26, 202555

Recently reported results: BIG-Bench Extra Hard Original — micro average

Latest version: BIG-Bench Extra Hard Original — micro average

ModelScoreSource
Gemma 4 E2B Instruct (thinking mode)21.9%huggingface.coMicro-average accuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 26B A4B Instruct (thinking mode)64.8%huggingface.coMicro-average accuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 12B Instruct (thinking mode)53.0%huggingface.coMicro-average accuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 E4B Instruct (thinking mode)33.1%huggingface.coMicro-average accuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord
Gemma 4 31B Instruct (thinking mode)74.4%huggingface.coMicro-average accuracyReported Model cardGoogleSelf-reportedGoogle DeepMindRecord