titleEmbeddingGemma 2: Unified and Efficient Multimodal Representations
authorsHenrique Schechter Vera, Sahil Dua, Jovan Andonov, Sara Smoot, Sebastian Russo, Ivan Llanos, Max Matveev, Jae Hyeon Yoo, Kat Black, Shih-Cheng Huang, Josiah Aklilu, Maarten Grootendorst, Omar Sanseviero, Chandramouli Amarnath, Ian Ballantyne, Zhe Li, Daniel Salz, Ibrahim Alabdulmohsin, Ilya Philippov, Kristen Quan, Glenn Cameron, Edouard Yvinec, Luciano Martins, Doug Reid, Sindhu Raghuram Panyam, Weiyi Wang, Hengxuan Ying, Madhuri Shanbhogue, Tanmaya Dabral, Shanfeng Zhang, Juyeong Ji, Renjie Wu, Joe Zou, Xu Chen, Yunan Zhang, Yu-Hui Chen, Matthew Soulanille, Vinod Mamillapalli, Yi-Chun Kuo, Wai Hon Law, Matt Kreileder, Alice Zheng, Suril Shah, Mogan Shieh, Pooja Consul, Alberto Montes, Andreas Hess, Andres Sevilla, Babak Samari, Brendan McCloskey, Brian Potetz, Daniel Cer, Georgios Kathareios, Gustavo Hernández Ábrego, Hesen Zhang, Ji Ma, Jiaxing Wu, Keegan Mosley, Paul Suganthan, Rui Meng, Shen Gao, Shengcao Cao, Yi-Ting Chen, Yichang Chen, Zhongli Ding, Kendra Byrne, Cormac Brick, Olivier Lacombe, Qin Yin, Tris Warkentin, Armand Joulin, Tom Duerig, Mojtaba Seyedhosseini
date2026-10-27
venuearXiv (preprint, not peer reviewed)
idarXiv:2610.37364
affiliationGoogle
upvotes30 (Hugging Face Daily Papers, 2026-11-01)
linksabs pdf hf
abstractWe 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.

cite

@misc{vera2026embeddinggemma,
  title={EmbeddingGemma 2: Unified and Efficient Multimodal Representations},
  author={Henrique Schechter Vera and Sahil Dua and Jovan Andonov and Sara Smoot and Sebastian Russo and Ivan Llanos and Max Matveev and Jae Hyeon Yoo and Kat Black and Shih-Cheng Huang and Josiah Aklilu and Maarten Grootendorst and Omar Sanseviero and Chandramouli Amarnath and Ian Ballantyne and Zhe Li and Daniel Salz and Ibrahim Alabdulmohsin and Ilya Philippov and Kristen Quan and Glenn Cameron and Edouard Yvinec and Luciano Martins and Doug Reid and Sindhu Raghuram Panyam and Weiyi Wang and Hengxuan Ying and Madhuri Shanbhogue and Tanmaya Dabral and Shanfeng Zhang and Juyeong Ji and Renjie Wu and Joe Zou and Xu Chen and Yunan Zhang and Yu-Hui Chen and Matthew Soulanille and Vinod Mamillapalli and Yi-Chun Kuo and Wai Hon Law and Matt Kreileder and Alice Zheng and Suril Shah and Mogan Shieh and Pooja Consul and Alberto Montes and Andreas Hess and Andres Sevilla and Babak Samari and Brendan McCloskey and Brian Potetz and Daniel Cer and Georgios Kathareios and Gustavo Hernández Ábrego and Hesen Zhang and Ji Ma and Jiaxing Wu and Keegan Mosley and Paul Suganthan and Rui Meng and Shen Gao and Shengcao Cao and Yi-Ting Chen and Yichang Chen and Zhongli Ding and Kendra Byrne and Cormac Brick and Olivier Lacombe and Qin Yin and Tris Warkentin and Armand Joulin and Tom Duerig and Mojtaba Seyedhosseini},
  year={2026},
  eprint={2610.37364},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2610.37364}
}