Google open-sources EmbeddingGemma 2: 740M multimodal embeddings that run on a Pixel
Prompt Engineering · youtube · 2026-10-09
Google released EmbeddingGemma 2, an open-source multimodal embedding model aimed at on-device search and RAG:
- Maps photos, videos, voice notes and text into one vector space, so a spoken query can retrieve the photo it describes without captioning or transcription steps
- 740M parameters under Apache 2.0; you can load just the 270M text portion
- Runs on a Pixel in under 600 MB of memory per Google
- Supports multimodal retrieval, classification and clustering
A full video with a free Colab notebook is available, and the model is on Hugging Face (google/embeddinggemma-2).
More from Models
- Leaked: X to bundle Grok and Cursor into XPass subscriptions from $8 to $200/month — alexcovo_eth · 2026-10-09
- Theorist Lance Fortnow: Claude Writes Math Better, Rewrites Unique Games Paper — fortnow · 2026-10-09
- Developer Hails Claude Opus 5.5 Plus 6.1 Sol as a Dream Coding Combo — himanshustwts · 2026-10-09
- ChatGPT removes model picker — which model is actually behind it now? — py-net · 2026-10-09
- minchoi Puts Grok Bot Through Its Paces: Handles Most Tasks, Learns From Misses — minchoi · 2026-10-09
- Anthropic's OSS Scanner: Claude models found 29,000 vulnerabilities, humans could review only 6,000 — npinto · 2026-10-09