Gemma 4 lands on phones, gets agents

This is a meaningful step for local AI: Google is making phone-class hardware feel like a real agent runtime, not just a demo surface. The bigger story is less “runs on phone” and more “mobile edge models can now execute workflows.”

• –Gemma 4’s E2B and E4B variants are tuned for low-memory, offline inference, which matters more than raw benchmark bragging for real-world mobile deployment
• –Agent Skills in AI Edge Gallery turn on-device models into something closer to programmable assistants, with skills for retrieval, summaries, visualizations, and multimodal composition
• –Support across iOS, Android, desktop, web, and edge hardware gives developers a single open model family to prototype across a wide deployment surface
• –The Apache 2.0 license makes this especially attractive for builders who want local control without model licensing friction
• –If the on-device stack keeps improving, expect more privacy-first, latency-sensitive apps to move logic off cloud APIs and onto the device