Nanbeige (南北阁, “South and North Pavilion Towers”) is an artificial intelligence research unit of Kanzhun Limited, a Beijing-based technology company that operates BOSS Zhipin (BOSS直聘), China’s largest online recruitment platform. The lab was founded in 2023 as an internal initiative to develop foundation models. The public name “Nanbeige” became established as the brand of the lab’s LLM research direction.
The first generation, Nanbeige-16B (2023–2024), was released in Base and Chat variants and with context up to 32,768 tokens; the model was trained on a corpus of 2.5 trillion tokens and distributed under Apache 2.0 with an additional license agreement. At the time, this was the standard scale for open models of that period, but even then the lab signaled its commitment to open weights.
A strategic shift occurred with the move to the Nanbeige4 family, built around a 3 billion parameter architecture. The Nanbeige4-3B technical report documents two key contributions: Fine-Grained Warmup-Stable-Decay (FG-WSD), a training scheduler that progressively refines data mixtures at different stages, and Dual Preference Distillation (DPD), a distillation method that combines generative refinement with reconstruction of reasoning chains. These mechanisms enabled a 3B-parameter model to compete with LLMs 10 times larger. This result defined Nanbeige’s research position: not a race for scale, but extracting maximum efficiency from a small number of parameters through data quality and multi-stage post-training.
The next step was the Looped Transformer architectural innovation in Nanbeige4.2-3B. Instead of increasing the number of parameters, the model uses the same stack of 22 layers twice in the forward pass (num_loops=2), which increased effective depth and capacity without adding weights: among open models of comparable size, reusing layers to increase effective depth remains a rare architectural choice. Here the company consistently develops the “small generalist” line: the model simultaneously demonstrates agentic behavior, code generation, and general reasoning, positioning itself as “a small general model that reasons, aligns, and acts.” Unlike narrowly specialized 3B models, Nanbeige4.1 claims the role of a single agentic core for tasks that traditionally require larger models.
The flagship release Nanbeige4.3-33B (October 24, 2026, Apache 2.0) marks a change in architectural paradigm: after the experimental Looped Transformer, the lab returned to a dense GQA architecture based on Llama models. The model was initially designed as an “agent teacher” for subsequent distillation into a 3B version, but it showed benchmark results above Qwen3.8-27B and Gemma4-31B and broke into the competitive TOP of the AI Intelligence Index.
Separate mention should be made of Depth Hybrid Attention (DHA), an architectural research effort by the lab that adapts hybrid attention schemes (Depth Window Attention for GQA+SWA; Depth Delta Attention for MLA+KDA) while maintaining a low KV-cache footprint. DHA is not yet part of the flagship releases, but it documents the search direction toward more efficient hybrid architectures.
Thus Nanbeige occupies a niche different from large labs competing through parameter scaling. Its track is focused on the efficiency of small models. In the global ecosystem, the lab acts simultaneously as an open-weight driver (all key models under Apache 2.0) and as a research unit of a corporate platform, which explains its applied focus on agentic programming and office automation.