Google DeepMind’s Omar Sanseviero announced on May 24 that Gemma 4 has surpassed 120 million downloads since its April 2 release under the Apache 2.0 license, calling the on-device AI trend “wild.” The figure adds to the broader Gemma family’s tally of over 400 million cumulative downloads across all generations and more than 100,000 community variants on Hugging Face. Gemma 4 launched in four sizes — E2B, E4B, a 26B Mixture-of-Experts, and a 31B dense model — with the two smallest variants co-developed with Qualcomm and MediaTek for efficient on-device inference, running in under 1.5 GB of memory at 4-bit quantization.
The milestone arrived alongside a field demo from Google engineers who took an experimental Gemma 4 app into a remote area with no cell service or Wi-Fi, testing offline visual understanding, handwriting-based math reasoning, and tool calling on Pixel hardware paired with prototype display glasses — all without a single server request. Google designed the E2B and E4B models as the foundation for Gemini Nano 4, Android’s next system-wide on-device model due on flagship devices later this year; Gemini Nano already runs on over 140 million Android devices. The combination of a multimodal architecture, native function-calling, and offline capability positions Gemma 4 as a direct platform play against Apple’s on-device intelligence stack.