EmbeddingGemma 2: Unified and Efficient Multimodal Representations
Google releases EmbeddingGemma 2, an open 744M-parameter embedding model that places text, images, audio, video and mixed inputs in one vector space and sets a new quality-for-size frontier on seven benchmark suites.
Ask this paper
Architecture. Modality encoders turn media into soft tokens for a shared transformer backbone, so encoders can be loaded selectively. The model runs at 271M parameters text-only, 439M as vision-language and 577M as audio-language.
Training. Simplified per-layer embeddings and a multi-stage curriculum that progresses through modalities and objectives keep quality high on unimodal, cross-modal and mixed-modal tasks.
Results. Across MTEB (multilingual, English, code), MIEB, MMEB-v2, MAEB and MSEB, it beats specialist models and models two to twelve times larger. It scores 61.36 on MTEB Multilingual v2 and 68.46 on MTEB English v2 in bfloat16.
Deployment options. It supports adjustable vision token budgets, truncatable embedding dimensions and quantized weights. The int4 version loses about 0.4 points on MTEB Multilingual. Weights are on Hugging Face and LiteRT.
Abstract
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.