Gemma 4 brings capable, multimodal AI directly to Android devices. In this session, we'll trace how Gemma evolved into a model family built for efficiency and explore what the edge-optimized E2B and E4B models can do on-device, from vision and audio understanding to agentic function calling. We'll cover the architecture behind that efficiency, including Per-Layer Embeddings (PLE), and show how Gemma and Gemini Nano work together through AICore. You'll learn how Quantization-Aware Training (QAT) preserves model capability within 1.1–3.5 GB RAM footprints, while Multi-Token Prediction (MTP) accelerates decoding speeds 1.5x–3x for responsive, real-time experiences.
Speakers: Sara Smoot
Chapters:
0:00 - Introduction and evolution of Gemma models
3:58 - Multimodal capabilities and benchmarks
9:45 - On-device hardware and model architectures
15:07 - Quantization techniques and optimization
18:35 - System integration and safety architecture
21:42 - Local on-device demos
25:27 - Resources and next steps
Watch the intelligent apps sessions from Android Dev Summit 2026 → https://goo.gle/ADS26-Intelligent-Apps
Watch all the sessions from Android Dev Summit 2026 → https://goo.gle/ADS26-All-Sessions
Subscribe to Android Developers → https://goo.gle/AndroidDevs
#AndroidDevSummit2026
Event: Android Dev Summit 2026
Products Mentioned: Gemini Nano, Gemma 4, Gemma for Android, AICore