2025: On‑Device AI Goes Mainstream on Android
summary
Edge AI already runs on phones: Gemma 3 4B ≈3.2 GB (4‑bit) proves small LLMs fit offline, paired with on‑device vector database / RAG for semantic search. Demo app indexes screenshots with OCR, image embeddings, ML Kit labeling and a quantized Gamma model for summaries and keyword extraction. Fine‑tuning TF Lite classifiers proves painful; emulator testing fails, real devices work. Resulting app returns OCR, semantic, category and image matches locally and shows latency and cost benefits compared to cloud calls.
AI-written summary · about
chapters (8)
- 0:00Introduction to edge AI
- 1:22Benefits of edge computing
- 3:27The hybrid AI paradigm
- 5:08Core technologies and costs
- 7:41Sustainability and usage
- 8:03Technical implementation
- 18:47Practical project challenges
- 18:59App development experience
related talks
24mAccelerating AI on Edge
31mBringing AI on-Device: From Cloud to Edge
41mBuild intelligent Android apps with Google's AI
55mEdgeAI - AI in the Palm of Your Hand
47mMyths and Reality of Android Machine Learning
43mWhat’s new in Android machine learning
10mAI in your pocket: Building intelligent Android apps
41mGen AI for Android Developers