Gemma Scope 2
Gemma Scope 2: open interpretability tools for every Gemma 3 size (270M to 27B), built from about 110 PB of data and 1T+ trained parameters.
Combines sparse autoencoders and transcoders so researchers can study jailbreaks, hallucinations and sycophancy inside the model. Google calls it the largest open-source interpretability release by a lab to date. Demo via Neuronpedia.
- Date
- Friday, 19 December 2025
- Lab
- Google DeepMind
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Training scale | ~110 PB data, >1T parameters across all released SAEs/transcoders | company |
Page is dated 2025-12-19; the DeepMind RSS feed lists 2025-12-16. "Largest" is Google's claim.
Sources
This record was checked against its sources on 6 October 2026. How we check