NVIDIA Releases Star Elastic, Bundling 30B, 23B, and 12B Models in One Checkpoint

NVIDIA researchers have proposed Star Elastic, a post-training method that embeds nested submodels at varying parameter budgets within a single parent reasoning model. [1] Applied to the Nemotron Nano v3 architecture, which features a hybrid Mamba, Transformer, and Mixture-of-Experts design, the technique yields 23-billion and 12-billion parameter variants from a 30-billion parameter parent checkpoint. [1] All three configurations are produced from a single training run of approximately 160 billion tokens and can be extracted without additional fine-tuning. [1] The extracted 23-billion parameter variant scores 85.63 on the AIME-2025 benchmark, outperforming the Qwen3-30B-A3B model's score of 80.00 despite utilizing fewer active parameters. [2]

The nested weight-sharing mechanism allows smaller submodels to reuse the most important weights from the parent model, ranking components like attention heads and MoE experts by their contribution to accuracy. [1] The research team reports that this method achieves a 360-times token reduction compared to pretraining each variant from scratch, as well as a seven-fold reduction over prior compression methods. [2] The 12-billion parameter variant operates at 2.4 times the throughput of the 30-billion parent on an H100 GPU at bfloat16. [2] The model was released on Hugging Face on May 7, 2026, under the Nvidia Open Model License for commercial use. [3]