Gemma Scope
400+ open JumpReLU sparse autoencoders (30M+ features) for Gemma 2 2B and 9B, a 'microscope' for interpretability research.
First large open suite of sparse autoencoders on every layer and sublayer of a frontier-lab open model, plus the Mishax tooling and Neuronpedia demos. Lowers the cost of mechanistic interpretability research.
- Date
- Wednesday, 31 July 2024
- Lab
- Google DeepMind
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SAEs released | 400+ (30M+ features) Gemma 2 2B and 9B | company |
Gemma Scope 2 followed in 2025-12.
Sources
- deepmind.google/blog/gemma-scope-helping-the-safety-community-shed-light-on-the-inner-work
- developers.googleblog.com/en/smaller-safer-more-transparent-advancing-responsible-ai-with-
This record was checked against its sources on 6 October 2026. How we check