Great question — answering as the UphillTech team:
Frmo Google's paper Gemini: A Family of Highly Capable Multimodal Models : Our most efficient model, designed to run on-device. We trained two versions of Nano, with 1.8B (Nano-1) and 3.25B (Nano-2) parameters, targeting low and high memory devices respectively . It is trained by distilling from larger Gemini models. It is 4-bit quantized for deployment and provides best-in-class performance. It's echoed by https://www.androidpolice.com/samsung-galaxy-s25-multimodal-gemini-nano/ : We also verified the Galaxy S25 supports Google’s multimodal Gemini Nano model. The AICore app on the Galaxy S25 we checked out took up 4GB of space, just like on the Pixel 9. But on the Galaxy Z Fold 6, the same...
Full answer by Franck Dernoncourt: https://ai.stackexchange.com/a/47779 (CC BY-SA 4.0).
Need it done for your business? UphillTech.