Google DeepMind Ships Gemma 4: Four Open Models (Apache 2.0) with Native Function-Calling, On-Device Agents, and Variable-Resolution Multimodal Input

Google Blog·medium signal

Google DeepMind released Gemma 4 in four sizes (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0, purpose-built for on-device agentic AI with native function-calling, structured JSON output, and system instructions — enabling autonomous agents that interact with tools and APIs without specialized fine-tuning. All models natively process video, images, and audio (E2B/E4B), with Gemma 4 serving as the base for Gemini Nano 4 on Android with up to 4x speed improvement and 60% less battery usage. NVIDIA confirmed RTX acceleration support, making local agentic AI on consumer hardware production-viable.