Model tensors plus live state
MiB
Lower is better
BF16 model tensors plus cache or state after a 2,048-token prefill at batch 1. Framework and temporary workspace memory are excluded.