EmbeddingGemma 2: an open, lightweight multimodal embedding model
Google’s new 740M-parameter embedder puts text, code, images, audio, and video into one local search space.
EmbeddingGemma 2 is released under Apache 2.0 and built on Gemma 4, with a modular setup that can run text-only at 270M parameters or add vision and audio encoders for full multimodal use. Google says it leads sub-1B multimodal embedding models on benchmarks including MTEB Code and MAEB, with a 9.92-point code gain over the prior EmbeddingGemma. The model supports an 8K token context window and can process local media combinations such as audio, images, and video frames. Google positions it for offline semantic search, routing, and on-device RAG, with weights available on Hugging Face and Kaggle. Google DeepMind's note
EmbeddingGemma 2 is released under Apache 2.0 and built on Gemma 4, with a modular setup that can run text-only at 270M parameters or add vision and audio encoders for full multimodal use. Google says it leads sub-1B multimodal embedding models on benchmarks including MTEB Code and MAEB, with a 9.92-point code gain over the prior EmbeddingGemma. The model supports an 8K token context window and can process local media combinations such as audio, images, and video frames. Google positions it for offline semantic search, routing, and on-device RAG, with weights available on Hugging Face and Kaggle. Google DeepMind's note
score 7