Gemma 4 Model Family
From mobile devices to developer workstations — a size optimized for every deployment target.
Gemma 4 E2B
Mobile · EdgeEngineered for maximum memory efficiency on edge devices. Activates an effective 2B parameter footprint to preserve RAM and battery life. Runs completely offline.
Gemma 4 E4B
Mobile · EdgeHigher capability edge model with audio and visual understanding. Integrates with Android AICore Developer Preview and ML Kit GenAI Prompt API for production use.
Gemma 4 26B MoE
Workstation · #6 OpenMixture-of-Experts model activating only 3.8B parameters during inference for exceptional tokens-per-second throughput. #6 open model on Arena AI leaderboard.
Gemma 4 31B Dense
Workstation · #3 OpenMaximum raw quality and capability. The premier model for fine-tuning and research. Fits on a single 80GB H100 GPU. Currently the #3 open model in the world on Arena AI text leaderboard.
Benchmark Performance
Evaluated against industry-standard datasets. See the full model card for additional benchmarks.
| Benchmark | Task | 31B Dense | 26B MoE | Gemma 3 27B (prior gen) |
|---|---|---|---|---|
| Arena AI (Text) | Human preference | 1452 | 1441 | 1365 |
| MMMLU | Multilingual Q&A | 85.2% | 82.6% | 67.6% |
| MMMU Pro | Multimodal reasoning | 76.9% | 73.8% | 49.7% |
| AIME 2026 | Mathematics | 89.2% | 88.3% | 20.8% |
| LiveCodeBench v6 | Competitive coding | 80.0% | 77.1% | 29.1% |
| GPQA Diamond | Scientific knowledge | 84.3% | 82.3% | 42.4% |
| τ2-bench Agentic | Agentic tool use | 86.4% | 85.5% | 6.6% |
Source: Google DeepMind · Arena AI scores as of April 2, 2026.
Massive context windows
Process entire codebases, books, or long documents in a single prompt.
Download Model Weights
Apache 2.0 license — free for commercial use.