BOSS Zhipin releases Nanbeige4.3-33B open model
The simple Llama-style architecture means almost any inference engine can run it out of the box, at the cost of high memory usage and no built-in speculative decoding beyond external MTP. Agent-focused teams looking for an open teacher model to distil from should benchmark it against Qwen3.8-27B now, and watch for the promised sub-10B distilled releases. BOSS Zhipin being a recruiter entering open-weight modelling is itself notable for the China AI ecosystem.
BOSS Zhipin's large language model lab Nanbeige (南北阁) released Nanbeige4.3-33B on HuggingFace, its second model after Nanbeige4.2 from August 2025. The 33B Dense version is optimised for Agent capabilities, scoring above Qwen3.8-27B on agent benchmarks, though its coding performance is weaker. Unlike the fully custom Looped Transformer of Nanbeige4.2, this model uses a standard Llama-style architecture with full attention and GQA, and supports 512K context by raising rope_theta to 120 million without dynamic YaRN or RoPE compression. The official positioning is a teacher model, with smaller distilled models expected later.
BOSS直聘刚发了个 Nanbeige4.3-33B
说实话我在HuggingFace上刚刷到的时候还以为是个欧洲模型, 硬是没读出来 Nanbeige 是啥, 搜了下发现其实是 BOSS 直聘的大语言模型实验室, 叫南北阁.
这次发布的 33B Dense 版本特别优化了模型的Agent能力, 我给大家做了个对比chart, 哪个面积大哪个就更强, 从Agent分数上来看, 甚至比Qwen3.8-27B 还要高. 当然其它的诸如写代码能力就差一些了.
它们之前8月份也发过 Nanbeige4.2, 是完全的自研架构 Looped Transformer, 而这次则用了非常保守的标准 llama 架构, 就是全注意力+GQA. 这也意味着, 这个模型推理起来还是很吃内存的. 以及只能用外置MTP加速了. 另外长上下文方案可能也比较糙, 模型上来就支持512K上下文, 但是没有配置动态 YaRN 或复杂的 RoPE 压缩算法. 就是直接把旋转位置编码的基数(rope_theta)调到了1.2亿来支持.
当然这么搞也是有好处的, 首先就是架构很简单, 几乎所有推理引擎都能跑, 其次官方的定位是教师模型, 所以估计后续会发布一些小的蒸馏模型供实际使用. (目前还没发). 等它们发布了我有空给大家来个10B以下小模型横评.
#Nanbeige #boss直聘
