I announced EmbeddingGemma 2, Google’s first open, natively multimodal embedding model designed for on‑device efficiency.
- model size = 740 M parameters
Performance claim: outperforms some specialist models more than twice its size.
- supported modalities = text, code, image, video, audio
- released open‑source on Hugging Face
- intended to pair with Gemma 4 for offline, privacy‑first RAG