Gemma 4 26B: benchmark results
Google's open-weight Gemma 4 26B, a flagship model of the Gemma 4 generation (April 2026). Provider: Google. Released 2026-04-02. Access: Open.
Aggregate Rating 1659 ± 17, rank #292 of 1656 rated models, from 63 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MMOOC - Missing Knowledge and Background (Multiple Choice) | 61.25 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| MMOOC - Missing Knowledge and Background (Open-Ended VQA) | 80 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| MMOOC - Multimodal Ambiguity (Multiple Choice) | 70.25 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| MMOOC - Multimodal Ambiguity (Open-Ended VQA) | 79.25 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| MMOOC - Unclear Logical and Symbolic (Multiple Choice) | 57.25 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| MMOOC - Unclear Logical and Symbolic (Open-Ended VQA) | 58 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| MMOOC - Visual False Premises (Open-Ended VQA) | 86.5 | Out-of-context score (%; mean of the refusal rate, the share | 100 |
| SEA-HELM | 74.81 | Mean Score (%) | 98.2 |
| MMOOC - Missing Knowledge and Background (Yes/No) | 76.25 | Out-of-context score (%; mean of the refusal rate, the share | 97.1 |
| MMOOC - Misleading Premise (Multiple Choice) | 89.25 | Shifted in-context score (%; mean of accuracy on the answera | 94.1 |
| MMOOC - Uncertain Spatial and Physical Context (Multiple Choice) | 85 | Out-of-context score (%; mean of the refusal rate, the share | 94.1 |
| MMOOC - Unclear Logical and Symbolic (Yes/No) | 49 | Out-of-context score (%; mean of the refusal rate, the share | 94.1 |
Interactive version: theaggregate.ai/model?slug=gemma-4-26b · How It Works · Data refreshed daily, snapshot 2026-10-20.