Reddit user rebuilds Gemma 4 31b into 26b version

A Reddit user is manually rebuilding Google's Gemma 4 31b model into a smaller 26b variant by removing selected SWA (sliding window attention) layers. The project involves running ablation tests to determine which SWA layers can be removed while maintaining performance. The goal is to create a more efficient version of the model for local use.