Gemma 4 12B: benchmark results
Google DeepMind's Gemma 4 12B, an encoder-free multimodal model handling text, image and audio input, with a 256K context window and support for 140+ languages. Provider: Google. Released 2026-06-03. Access: Open.
Aggregate Rating 1573 ± 16, rank #535 of 1656 rated models, from 105 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA IFBench | 73.54 | Accuracy (%) | 91.3 |
| AI for Education Pedagogy - Technology | 85.85 | Accuracy (%) | 88.5 |
| SEA-HELM | 69.31 | Mean Score (%) | 82.5 |
| LinguaGuard — original wording balanced accuracy — structured step prompt | 0.85 | Balanced accuracy | 76.9 |
| LLM Stats (MathVision) | 79.7 | Score (%) | 70 |
| AA Omniscience - Software Engineering (SWE) - HTML | 34.69 | Accuracy (%) | 67.7 |
| IHBench - Recovery Quality - Filler | 25 | Recovery pass rate (%; 60 backchannel fillers, where the tur | 61.5 |
| ECGQuest | 62.9 | Accuracy (%; 1,050 held-out true/false questions on ECG know | 61.4 |
| AA Humanity's Last Exam | 15.66 | Accuracy (%) | 61 |
| IHBench - Task Fulfillment | 51.1 | Win rate against GPT-4o Audio (%; share of 428 interruption | 60 |
| AA Omniscience - Software Engineering (SWE) - Rust | 50 | Accuracy (%) | 59 |
| AA-Omniscience Hallucination Rate | 73.41 | Hallucination Rate (%) | 58.8 |
Interactive version: theaggregate.ai/model?slug=gemma-4-12b · How It Works · Data refreshed daily, snapshot 2026-10-20.