Gemma 4 31B (IT) (Thinking): benchmark results
Provider: Google. Released 2026-04-02. Access: Open.
Aggregate Rating 1708 ± 10, rank #335 of 2170 rated models, from 149 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FrameBench - Japanese | 99.1 | Accuracy (%; mean of five prompt templates) | 100 |
| IndicBankBench - Long Context | 60.67 | Strict pass^3 (%; 89 cases passed out of the 89 cases of the | 100 |
| NitiBench+ - Cited sections precision | 0.64 | Precision (0–1) | 100 |
| PostEDA-Bench (ReAct) - PPA-Mono | 64.56 | Success rate (%; 35 single-objective PPA tasks, mean of 5 te | 100 |
| RedHat Gemma4 FP8 - Thinking - BFCLv4 Agentic | 65.82 | Native accuracy (%) | 100 |
| RedHat Gemma4 FP8 - Thinking - BFCLv4 Overall | 70.77 | Native accuracy (%) | 100 |
| RedHat Gemma4 FP8 - Thinking - GPQA Diamond (0-shot, pass@1) | 86.36 | Native accuracy (%) | 100 |
| Swallow - Post-trained English - Average | 87.4 | Average Score (%) | 100 |
| Swallow - Post-trained English - IFBench | 76.5 | Instruction-Level Strict Accuracy (%) | 100 |
| Swallow - Post-trained English - LiveCodeBench | 81 | Pass@1 (%) | 100 |
| Swallow - Post-trained Japanese - WMT20 En-Ja | 28.5 | BLEU | 100 |
| Werewolf Accusation Benchmark - Role Separation | 1.69 | Belief separation (wolves minus villagers, -3..+3 scale) | 100 |
Interactive version: theaggregate.ai/model?slug=gemma-4-31b-it-thinking · How It Works · Data refreshed daily, snapshot 2026-11-05.