Mini toolsNanbeige4.2-3B
Nanbeige4.2-3B is a compact agentic language model with only 3B non-embedding parameters, built using a Looped Transformer architecture that reuses transformer layers to increase capacity without adding parameters. It was trained with diverse supervised fine-tuning across real-world and synthesized agent environments, followed by reinforcement learning combining outcome and process rewards for training stability. The model supports configurable reasoning (thinking) modes, multi-turn tool use with XML or JSON tool-call formats, and strong performance on tool-use, office-agent, and code-agent benchmarks such as SWE-Bench and Terminal-Bench, often outperforming larger models. It also shows strong mathematical, scientific, and coding reasoning. Designed to be lightweight enough for local deployment, it can function as a local personal assistant supporting daily assistance, office workflows, and deep research tasks when paired with an agentic scaffold.
New
Text
Overview
Nanbeige4.2-3B is a compact 3B-parameter agentic language model using a Looped Transformer architecture that reuses transformer layers to boost capacity without adding parameters. It combines strong tool-use and coding-agent performance with broad reasoning across math, science, and general agent benchmarks, supporting configurable thinking and tool-calling modes for multi-turn agentic workflows.
About Nanbeige LLM Lab
Nanbeige (Chinese: ๅๅ้, which means "South and North Pavilion") is a large language model (LLM) research team from BOSS Zhipin.
Location:
Bejing, CN
Last updated: July 22, 2026
