FrontierMay 9, 2026via MarkTechPost
NVIDIA AI Releases Star Elastic: One Checkpoint that Contains 30B, 23B, and 12B Reasoning Models with Zero-Shot Slicing
Why it matters
NVIDIA demonstrates a fundamental shift in model scaling efficiency: instead of training separate variants, one post-training method now produces multiple nested models from a single run. This directly impacts inference cost and deployment flexibility for enterprises choosing between model sizes.
Key signals
- Star Elastic enables 30B, 23B, and 12B models in single checkpoint via zero-shot slicing
- 360× token reduction vs. training each model separately (trained on 160B tokens total)
- Elastic budget control achieves 16% higher accuracy and 1.9× lower latency vs. standard budget control
- Nested FP8 and NVFP4 checkpoints bring full model family within reach of RTX-class GPUs
- Built on Nemotron Elastic framework, applied to Nemotron Nano v3
- Post-training method, not architecture change
The hook
360× token reduction. NVIDIA's Star Elastic embeds three reasoning models in one checkpoint—30B, 23B, and 12B—eliminating redundant training entirely.
NVIDIA researchers have introduced Star Elastic, a post-training method that embeds multiple nested reasoning models — at 30B, 23B, and 12B parameter scales — inside a single checkpoint, eliminating the need for separate training runs or stored model weights per variant. Built on the Nemotron Elasti…