“NVIDIA's Nemotron-3 model found an effective ratio between standard attention layers for global information access and compressed state layers for efficiency.”