Google added a new Gemma model positioned as “differentially private,” alongside recent smaller releases. For teams training or deploying LLMs on sensitive corpora, the key question is whether this meaningfully reduces memorization risk enough to change what you can safely build and ship.
VaultGemma arrives as Google’s “differentially private” Gemma option
Google has released VaultGemma, a 1B-parameter model in the Gemma family, described as differentially private, according to the Gemma releases documentation. The release date listed is Sept. 13, 2025.
The VaultGemma launch follows two other recent Gemma updates also listed in the same release notes: EmbeddingGemma (308M parameters, Sept. 4, 2025) and Gemma 3 (270M parameters, Aug. 14, 2025). Together, the sequence signals Google is expanding the Gemma line across both capability (a 1B model) and workflow specialization (embeddings, and now a privacy-positioned model).
- Privacy posture may shift from policy-only to technical controls. If VaultGemma’s “differentially private” claim holds in practice for your use case, it can reduce the chance that sensitive training examples are reproduced in outputs—useful when internal data includes regulated or contractual constraints.
- Compliance teams will ask for specifics. “Differentially private” is not a single setting; data leaders should expect to document what DP guarantees apply (and where they don’t), and how that interacts with retention, access controls, and incident response for model outputs.
- Model choice becomes part of your data governance design. With Gemma variants spanning embeddings and DP-positioning, teams can start mapping model selection to data classification tiers (e.g., public vs. restricted) rather than treating model choice as purely an accuracy/latency decision.