Google EmbeddingGemma 2: Best On-Device Multimodal Embeddings (Here's Why)

Google just released EmbeddingGemma 2, a free embedding model that turns your photos, videos, audio and text into numbers that capture what they mean, so you can search just by describing what you want. The full version needs about 567 MB of memory, so it runs right on your phone. I built an app with it called ReelFinder that searches my own YouTube b-roll clips from a voice note. No typing, no captions, no tags. I tried 7 voice notes, including one in Hindi, and 5 of them found the right clip on the first try. You'll see the hits and the miss. In this video: What an embedding model is, explained as a "map of meaning" How search with it works in three steps Google's own demos: live camera photo search and the 47 ms dino race The five things Google calls out: best in its class for its size, modular blocks, smaller storage, low memory on a phone, and 4x more input at once My ReelFinder demo, hits and misses Who should actually use this (sensitive files, big media libraries) 🔗 LINKS Google announcement: https://blog.google/innovation-and-ai/technology/developers-tools/embeddinggemma-2/ ⏱️ CHAPTERS 00:00 What EmbeddingGemma 2 does 00:34 ReelFinder: searching my clips by voice 01:22 Embeddings explained: a map of meaning 02:01 Search in three steps 02:23 Google's demos 02:50 What's new since the first EmbeddingGemma 03:16 Benchmarks: best in its class for its size 03:45 Modular building blocks 04:05 Smaller storage 04:31 Memory on a phone 04:57 4x more input at once 05:13 Back to ReelFinder: indexing my library 05:31 More voice searches (and one miss) 06:18 Who should use this 06:47 Wrap-up I'm Surya, 20 years in enterprise tech. I take the cutting edge of AI and make it make sense. Subscribe for more. All opinions are my own and do not belong to my employer. #EmbeddingGemma #Gemma #GoogleAI #OnDeviceAI #AI