NVIDIA Releases Star Elastic: Single Checkpoint Contains Multiple Reasoning Models

NVIDIA released Star Elastic on May 9, 2026, a post-training method that embeds multiple nested reasoning models at 30B, 23B, and 12B parameter scales inside a single checkpoint. The method trains all three model variants in a single 160B-token run, achieving a 360× token reduction compared to pretraining each model from scratch. Star Elastic enables zero-shot model slicing without retraining and includes nested FP8 and NVFP4 checkpoints accessible on RTX-class GPUs.

Star Elastic training method

• ▪Star Elastic achieves a 360× token reduction compared to pretraining each model from scratch
• ▪Star Elastic was applied to Nemotron Nano v3
• ▪Star Elastic trains all three model variants in a single 160B-token run
• ▪Star Elastic is built on the Nemotron Elastic framework
• ▪Star Elastic is a post-training method introduced by NVIDIA researchers
• ▪NVIDIA released Star Elastic on May 9, 2026

Nested model architecture

• ▪Star Elastic embeds multiple nested reasoning models at 30B, 23B, and 12B parameter scales inside a single checkpoint
• ▪Star Elastic eliminates the need for separate training runs or stored model weights per variant

Elastic budget control

• ▪Elastic budget control delivers up to 16% higher accuracy compared to standard budget control
• ▪Elastic budget control is a novel inference scheme introduced by Star Elastic that uses a smaller submodel for the thinking phase and the full model for the final answer

Quantized checkpoint formats

• ▪Star Elastic's nested FP8 and NVFP4 checkpoints bring the full model family within reach of RTX-class GPUs
• ▪Star Elastic includes nested FP8 checkpoints