Safety
59d agoQuantization Effects On Gemma 3 Interpretability
Weight compression tests on Gemma 3 4B and 12B models reveal how 8-bit and 4-bit quantization impact model performance. This analysis compares compressed versions against uncompressed controls to map representational geometry. The findings clarify how reducing precision alters feature activation.
The Signal
Practitioners can now better predict how quantization degrades interpretability in smaller DeepMind models.