NVIDIA Star Elastic packs 30B, 23B, and 12B reasoning models into one checkpoint with 360× lower training cost.