We introduce VaultGemma 1B, a 1 billion parameter model within the Gemma family, fully trained with differential privacy. Pretrained on the identical data mixture used for the Gemma 2 series, VaultGemma 1B represents a significant step forward in privacy-preserving large language models. We openly release this model to the community
Cite
@article{arxiv.2510.15001,
title = {VaultGemma: A Differentially Private Gemma Model},
author = {Amer Sinha and Thomas Mesnard and Ryan McKenna and Daogao Liu and Christopher A. Choquette-Choo and Yangsibo Huang and Da Yu and George Kaissis and Zachary Charles and Ruibo Liu and Lynn Chua and Pritish Kamath and Pasin Manurangsi and Steve He and Chiyuan Zhang and Badih Ghazi and Borja De Balle Pigem and Prem Eruvbetine and Tris Warkentin and Armand Joulin and Ravi Kumar},
journal= {arXiv preprint arXiv:2510.15001},
year = {2025}
}