Descriptions:
Paige Bailey, engineering lead for developer relations at Google DeepMind, delivers a conference talk at AI Dev 26 in San Francisco covering Google’s current model portfolio and its real-world deployments. The centerpiece is Gemma 4, Google’s new open model family released under an Apache 2.0 license, available in four sizes: 2 billion parameters (mobile-ready), 4 billion (runs locally on a laptop), a 26 billion parameter mixture-of-experts model, and a 31 billion parameter dense model. Bailey notes Gemma 4 benchmarks favorably against models such as Claude Sonnet at comparable parameter counts and is available via Hugging Face.
The talk highlights several applied domains. On robotics, Gemini models power manipulation tasks on Google’s Mountain View campus and run on Stanford’s fully 3D-printable Pepper robots via Raspberry Pi, using Gemini for live conversation and vision understanding. For augmented reality, Gemini integrates with Meta Ray-Ban glasses and third-party smart glasses through a REST API, enabling real-time navigation, visual explanation, and AR experiences. Real-time speech translation — currently API-only — lets users speak in one language and hear a response in their own.
Bailey also covers Gemini 3.1 Live’s native audio token output and a fully on-device pipeline combining Gemma for speech-to-text, inference, and text-to-speech that keeps user data local. Developers can explore all capabilities through Google AI Studio, with a dedicated hands-on workshop session running the same day.
📺 Source: DeepLearningAI · Published May 20, 2026
🏷️ Format: Keynote Launch
