Google launches EmbeddingGemma 2, a 740M open multimodal embedding model for text, code, images, video and audio on-device

What changed

Google DeepMind launched EmbeddingGemma 2 on October 6. Released under Apache 2.0 and based on Gemma 4, the 740M-parameter model maps text, code, images, video and audio into a unified 768-dimensional embedding space. Its modular architecture can load about 270M parameters for text/code-only workloads and add vision or audio encoders when needed. Google gives quantized Pixel 11 Pro active-memory examples of roughly 191MB for text-only weights and 567MB for the full multimodal model, with an 8K context window for on-device search, RAG, media retrieval and routing. Google's benchmark-leadership claims are vendor evaluations rather than independent rankings.