The future of AI is moving on-device, bringing intelligence closer to users, faster, and much more private. With open-source language models and efficient runtimes, Android developers can now run AI directly on smartphones. This removes cloud dependency and gives users full control of their data.
In this talk, I will share how to build privacy-first Android apps using open-source AI models, covering the process from model selection to integration.
We will explore how to choose lightweight models such as Gemma 2B, Phi-3 Mini, or Gemini Nano that are suitable for mobile devices. I will also explain how to fine-tune these models with your own dataset using Hugging Face Transformers and QLoRA, and how to optimize them for Android using TensorFlow Lite and ONNX Runtime Mobile.
We will also explore on-device AI models through Android’s AI Core, showing how developers can leverage upcoming native support for running local models efficiently and securely on modern Android devices.
The session will include a live demo of an Android app running an open-source model locally, performing tasks like note assistance, summarization, or translation without sending any information outside the device.
This talk reflects the spirit of Free and Open Source Software by helping developers create intelligent, ethical, and privacy-focused mobile applications using transparent and community-driven technologies.