← Back to feed

NVIDIA released Star Elastic on May 7, a training technique that bakes multiple smaller models inside a single parent checkpoint. The approach, applied to Nemotron Nano v3, lets operators slice out 23B or 12B reasoning variants from a 30B parent without retraining.

How It Works

The parent model is Nemotron Nano v3: a hybrid Mamba2–Transformer–MoE architecture with 30B total parameters and 3.6B active parameters. Star Elastic trains this model with approximately 160B tokens while simultaneously learning to embed two nested submodels:

• 23B (2.8B active) — midpoint deployment tier
• 12B (nested inside the 23B) — edge and cost-optimized deployment

Submodels reuse the most important weights from the parent, identified through Router-Weighted Expert Activation Pruning (REAP) — a method that ranks experts by routing gate values combined with output magnitudes. The alternative (naive frequency-based pruning) is what most prior work uses; REAP provides a more principled signal aligned with actual contribution to output quality.

The training loss combines three terms: knowledge distillation from the parent, a router loss that penalizes deviation from the target resource budget, and task-specific objectives. The router learns architecture choices that improve accuracy rather than minimising a proxy metric.

Deployment Numbers

• 1.9x lower latency versus standard single-model deployment at equivalent parameter count
• Released under the NVIDIA Open Model License on HuggingFace on May 7, 2026
• Ready for commercial use

One current limitation: switching submodels mid-reasoning is not yet supported in standard vLLM. Operators commit to a submodel tier at deployment initialization.

The Argument for Nested Models

The standard deployment workflow for teams that need multiple model tiers involves separate training runs, separate checkpoints, and separate infrastructure — each model family requiring duplicated QA, serving infrastructure, and update pipelines. Star Elastic compresses that into one model family checkpoint: a single artifact that ships 30B for the reasoning workload, 12B for the latency-sensitive workload, and 23B for whatever sits in between.

This is not purely a research result. Nemotron Nano v3 ships as a production model, and NVIDIA positions Star Elastic as the architecture for how it manages multi-tier model families going forward.

Context

NVIDIA Nemotron 3 Nano Omni — the multimodal version of the same model family — shipped earlier this year and topped six leaderboards at 30B total / 3B active parameters. Star Elastic is the reasoning-optimised complement: same base architecture, new training methodology, three deployment sizes.

The approach echoes recent work from Microsoft (Phi-3-small/medium/large sharing a training lineage) and Google (Gemma 4 across four parameter tiers), but Star Elastic is the first system to produce the nested hierarchy from a single training run rather than a family of separate runs.