EmbeddingGemma 2: Unified and Efficient Multimodal Representations
EmbeddingGemma 2 puts text, images, audio, and video into one embedding space. It sets a new Pareto frontier, beating models two to twelve times larger.
The claim
One open model embeds text, images, audio, and video in one space, beating larger rivals.
The problem
Embedding models usually cover a single modality at a time.
Teams juggle separate specialist models for text, images, audio, and video.
Small devices cannot carry every full encoder for all modalities.
How it works
EmbeddingGemma 2 natively represents five input types and their combinations in one space.
Simplified per-layer embeddings feed the shared representation for every modality.
A multi-stage curriculum progresses modality first, then training objective.
Encoders load selectively for only the modalities that a device needs.
Vision token budgets, embedding dimensions, and quantized weights all stay adjustable.
Figures
What changes
The model sets a new Pareto frontier across all embedding benchmarks.
It outperforms specialist models and models two to twelve times larger.
Quantized weights and truncatable dimensions support deployment on small devices.
Limits
Benchmark details stay in the paper, not in the abstract.
Quality for each single modality is not broken out in the abstract.
The paper reports no latency, throughput, or memory figures.
Source · arXiv