Google announced Gemma 3n preview on May 20, 2025 as an open model built on the same shared architecture that will also power the next generation of Gemini Nano.
Google says Gemma 3n is optimized for everyday devices including phones, laptops and tablets, and the model handles text, visual data and audio input while generating text output.
Google's developer blog says Gemma 3n uses Per-Layer Embeddings, KVC sharing and advanced activation quantization to reduce the memory footprint.
Google's Kaggle model card says Gemma 3n uses selective parameter activation so the models can operate at effective 2B and 4B parameter sizes, below the total number of parameters they contain.
The Hugging Face model card for google/gemma-3n-E4B-it says the repository is the launch version of the instruction-tuned Gemma 3n E4B model for Transformers, with text, audio, vision image and video inputs.
The same model card says the raw parameter count of the E4B model is 8B, while the architecture lets it run with a memory footprint comparable to a traditional 4B model by offloading low-utilization matrices from the accelerator.
Google's developer blog says the preview was available through Google AI Studio for browser testing and through Google AI Edge for local text and image development at launch.
For local-app developers, the practical claim is narrow: Google is offering open weights and tooling for device-side multimodal experiments, while the cited pages do not provide an independent benchmark of real app latency or battery use.
