Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    APICircuit TracerNEWSteerSAE EvalsExportsSlackBlogPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlogGitHubSlackTwitterContact
    Gemma Scope 2: Comprehensive Suite of SAEs and Transcoders for Gemma 3
    Language Model Interpretability Team, Google DeepMind· December 2025
    Tutorial Notebook
    HuggingFace
    DeepMind Blog
    ⚠️ Rolling Release
    An initial release of artifacts and datasets are available now, with all artifacts expected to be finalized by January 16, 2026.
    Some data may be replaced or updated during this final verification and fine-tuning process. Please check the HuggingFace for details.
    Gemma Scope 2
    Demo
    Examining Safety-Relevant Features and Circuits in Gemma 3
    👋 New Here?
    If you're new to interpretability (the science of understanding what happens inside AI), we recommend you start with the original "Exploring Gemma Scope", which has more beginner-friendly interactive demos and content.
    This Gemma Scope 2 demo focuses on exploring safety-relevant features in Gemma 3 27B-IT, the largest model in the new Gemma 3 model series. Since the Gemma Scope 2 release also includes transcoders, cross-layer transcoders, and crosscoders, Neuronpedia is also adding support for circuit tracing with those new artifacts.
    🔢 Sections
    Do More
    Tutorial Notebook
    A Colab notebook for loading and using Gemma Scope 2 artifacts.
    Hugging Face
    Download the Gemma Scope 2 SAEs and transcoders.
    DeepMind Blog
    Read the official blog post about Gemma Scope 2.