Nanbeige4.2-3B Released: A 3B Parameter Model Using Looped Transformer Architecture

The Nanbeige LLM Lab has released Nanbeige4.2-3B, a compact 3B parameter agentic language model that utilizes a Looped Transformer architecture. This architecture reuses transformer layers iteratively to increase model capacity without increasing the overall parameter count. This release demonstrates that compact models can achieve high performance on agentic and coding tasks, challenging the need for larger models by outperforming models four times their size. It highlights the potential of parameter-efficient, recursion-based architectures in practical AI applications. Built on Nanbeige4.2-3B-Base, the model features only 3 billion non-embedding parameters. It is designed to combine strong agentic behavior with broad reasoning and alignment capabilities.