DiffusionGemma 26B-A4B: benchmark results
Google DeepMind's open-weight text-diffusion model (June 2026) built on the Gemma 4 26B-A4B MoE backbone: 25.2B total / 3.8B active parameters, denoising 256-token canvases in parallel for ~4x faster generation. Provider: Google. Released 2026-06-10. Access: Open.
Unified ELO 1562 ± 36, rank #784 of 2656 rated models, from 56 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA IFBench | 59.46 | Accuracy (%) | 70.9 |
| AA Omniscience - Software Engineering (SWE) - Julia | 12.5 | Accuracy (%) | 63.8 |
| AA Omniscience - Software Engineering (SWE) - Rust | 50 | Accuracy (%) | 59 |
| MedXpertQA | 49 | Score (self-reported) | 56.7 |
| AA Omniscience - Software Engineering (SWE) - Go | 18 | Accuracy (%) | 56.5 |
| AA Humanity's Last Exam | 10.84 | Accuracy (%) | 55.3 |
| AA Omniscience - Software Engineering (SWE) - PHP | 22 | Accuracy (%) | 55.3 |
| AA CritPt | 0.29 | Accuracy (%) | 52.3 |
| AA Omniscience - Software Engineering (SWE) - JavaScript | 27.27 | Accuracy (%) | 48.3 |
| ZeroEval GPQA Diamond | 73.2 | GPQA Diamond Score | 47.3 |
| AA-Omniscience Index - Software Engineering (SWE) - Go | -46 | Omniscience Index | 45.5 |
| AA Omniscience - Software Engineering (SWE) - Python | 16.08 | Accuracy (%) | 45.2 |
Interactive version: theaggregate.ai/model?slug=diffusiongemma-26b-a4b · How It Works · Data refreshed daily, snapshot 2026-09-19.