Google DeepMind has opened its latest interpretability toolkit, Gemma Scope 2, to the research community. The suite covers the entire Gemma 3 family, from 270 million to 27 billion parameters, and lets analysts step inside the model’s logic. By providing a full‑scale microscope, DeepMind aims to expose hidden risks and emergent behaviors that surface only at larger sizes.

Creating Gemma Scope 2 required storing roughly 110 Petabytes of traces and training over 1 trillion total parameters. The toolset blends sparse autoencoders and transcoders, trained on every layer of Gemma 3. This architecture enables researchers to dissect multi‑step computations, align internal states with reported reasoning, and scrutinize jailbreaks, refusals, and chain‑of‑thought faithfulness in chat‑tuned variants.

DeepMind positions Gemma Scope 2 as the largest open‑source interpretability release by an AI lab, inviting the community to debug emergent behaviors and accelerate safety interventions against hallucinations, sycophancy, and jailbreaks. An interactive demo hosted by Neuronpedia lets users experiment live, while the underlying Matryoshka training technique promises more precise concept detection than its predecessor.

By exposing the internal dynamics of Gemma 3, researchers can map how complex reasoning emerges and pinpoint failure modes before deployment. The toolkit’s open nature encourages cross‑lab collaboration, potentially standardizing interpretability benchmarks and fostering safer, more trustworthy language models for commercial and scientific use.