#gemma

Picking the model for the hardware I have

Picking a model for the trusted-AI build on 24 GB. The bench beat intuition: a MoE beat a dense model twice its size; 'thinking' only got in the way.