Nanbeige4.2-3B uses a looped transformer to expand capacity without more parameters

A research post presents Nanbeige4.2-3B as a compact agentic model pretrained on 28 trillion tokens with a layer-reusing Looped Transformer.