Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control

Training a family of large language models (LLMs), either from scratch or via iterative compression, is prohibitively expensive and inefficient, requiring separate training runs for each model in the family. In this paper, we introduce Star Elastic, a novel LLM post-training method that adds N nested submodels to a given parent reasoning model using the compute of one run (Nx savings) via a single post-training job. Beyond reducing training costs, Star Elastic also addresses a fundamental limitation in efficient reasoning: the rigidity of static architectures, which forces the allocation of constant resources regardless of token difficulty. By unlocking elastic budget control, Star Elastic enables a novel approach that uses different submodels for each reasoning phase (thinking and answering). Star Elastic supports (1) nesting along the SSM, embedding channel, MoE and FFN axes, (2) learning nested submodels via an end-to-end trainable router, and (3) curriculum-based knowledge distillation. We apply Star Elastic to the NVIDIA Nemotron Nano models; in particular, we demonstrate its effectiveness on hybrid MoE architectures with Nemotron Nano v3 (30B/3.6A), generating 23B (2.8A) and 12B (2.0A) variants with 160B training tokens. For Nemotron Nano v2 (12B), we produce 9B and 6B nested models using only 110B training tokens, achieving a 360x reduction versus training from scratch and a 7x reduction over state-of-the-art compression methods. All nested models match or outperform independently trained baselines of comparable size. Crucially, elastic budget control advances the accuracy--latency Pareto frontier, achieving up to 16% higher accuracy and 1.9x lower latency via dynamic per-phase model selection.

Large language models are often released in several sizes, so people can choose between higher accuracy and lower cost. Today, creating that family usually means training or compressing each size separately, which is expensive and wasteful. This work introduces Star Elastic, a method that turns one trained reasoning model into several smaller models nested inside it, like a set of measuring cups that fit together. The same model family can then choose how much computing power to use at different moments: for example, using a smaller model while working through a problem and a larger one when producing the final answer. In tests on NVIDIA’s Nemotron Nano 3 models, Star Elastic produced smaller 23B and 12B versions from a 30B parent model in one training run, while keeping strong accuracy. It also improved the speed–accuracy trade-off, reaching up to higher accuracy with lower latency. This could make capable reasoning AI cheaper to train, store, and run across different hardware settings.