[mmoskal]

Also ~noone runs h100 at home, ie at batch size 1. What matters is throughput. With 37b active parameters and a massive deployment throughout (per gpu) should be similar to Gemma.