The new reference architecture, developed with NVIDIA, provides a fully open stack for enterprises to own and customize, enabling full visibility into the model, harness, and runtime layers. This blueprint achieves benchmark-leading performance and over 10x lower inference costs.