SciCode Benchmark Results & Scores | Benchmark Registry

Versions of SciCode
Version	Released	Models	Results
SciCode OriginalLatest	July 18, 2024	9	9

Latest version: SciCode Original

Recently reported results for SciCode Original
Model	Score	Source
Gemma 4 E2B Instruct (thinking mode)	21.0%	arxiv.org for Gemma 4 E2B Instruct (thinking mode) on SciCode Original (opens in a new tab)AccuracyReported July 24, 2026Technical reportGoogle DeepMindSelf-reportedRecord for Gemma 4 E2B Instruct on SciCode Original
Gemma 4 26B A4B Instruct (thinking mode)	40.0%	arxiv.org for Gemma 4 26B A4B Instruct (thinking mode) on SciCode Original (opens in a new tab)AccuracyReported July 24, 2026Technical reportGoogle DeepMindSelf-reportedRecord for Gemma 4 26B A4B Instruct on SciCode Original
Gemma 4 12B Instruct (thinking mode)	38.0%	arxiv.org for Gemma 4 12B Instruct (thinking mode) on SciCode Original (opens in a new tab)AccuracyReported July 24, 2026Technical reportGoogle DeepMindSelf-reportedRecord for Gemma 4 12B Instruct on SciCode Original
Gemma 4 E4B Instruct (thinking mode)	24.0%	arxiv.org for Gemma 4 E4B Instruct (thinking mode) on SciCode Original (opens in a new tab)AccuracyReported July 24, 2026Technical reportGoogle DeepMindSelf-reportedRecord for Gemma 4 E4B Instruct on SciCode Original
Gemma 4 31B Instruct (thinking mode)	43.0%	arxiv.org for Gemma 4 31B Instruct (thinking mode) on SciCode Original (opens in a new tab)AccuracyReported July 24, 2026Technical reportGoogle DeepMindSelf-reportedRecord for Gemma 4 31B Instruct on SciCode Original