NVIDIA AI Releases Star Elastic: One Checkpoint That Contains 30B, 23B, and 12B Reasoning Models with Zero-Shot Slicing

NVIDIA researchers introduced Star Elastic, a post-training framework that learns nested submodels inside one parent model.

Applied to Nemotron Nano v3, it produces 30B, 23B, and 12B variants from a single checkpoint with shared weights and a learned router.

The approach aims to cut storage and serving costs by letting teams deploy multiple budgeted model sizes without extra fine-tuning.