#model-architecture

NVIDIA AI has released Star Elastic, an innovative architecture that packs 30B, 23B, and 12B reasoning models into a single checkpoint, enabling zero-shot slicing to dynamically switch between model scales without separate downloads.

A LocalLLaMA explainer argues that Gemma 4 E2B/E4B gain their efficiency from Per-Layer Embeddings. The key point is that many of those parameters behave more like large token lookup tables than always-active compute-heavy layers, which changes the inference trade-off.