Archive · 16 Sep 2025
Google Launches VaultGemma: A Privacy-Focused AI Model
Imagine teaching a super-smart computer, but you don't want it to accidentally share secrets it learned.
Google made a new AI called VaultGemma that is extra careful with private info.
They used a special method called differential privacy, which adds some 'noise' to the lessons so the computer doesn't remember exact facts.
It's like learning without memorizing everything.
The model is pretty big, but it still performed well.
This way, Google is working to keep data private, protecting user information.
Google developed VaultGemma, an AI model with privacy-preserving techniques.
VaultGemma uses differential privacy (DP) to protect training data.
The model has one billion parameters and is the largest DP-trained language model.
VaultGemma's weights are freely available for download.
Google researchers developed new scaling laws for DP to balance compute, privacy, and utility.
- Who
- What
- Launched VaultGemma, a new AI model.
- When
- September 12 (blog post)
- Where
- N/A
- Why
- To build a powerful and private AI by design.
- Model Name
- VaultGemma
- Developer
- Key Feature
- Privacy-preserving techniques
- Parameter Count
- One billion
- Training Method
- Differential Privacy (DP)
- Model Availability
- Free download from Hugging Face and Kaggle