Comment on: Google Gemma 4

curious about the "low compute overhead" claim - are you seeing meaningful performance gains over Llama models in the same parameter range? we're always evaluating new models for healthcare applications where inference speed matters a lot.