The Nanbeige team at Kanzhun, the company behind China's online recruitment platform BOSS Zhipin, has open-sourced Nanbeige4.3-33B, a model built for agentic coding and professional office tasks. The weights are available on the team's official Hugging Face repository under the Apache 2.0 license, which allows commercial use.
The team describes Nanbeige4.3 as a first step toward pushing the limits of lightweight agents. Its strategy is to build a high-capacity teacher agent first, then distill its agent skills into smaller models. Nanbeige4.3-33B is the flagship teacher for the series, and the team names a planned Nanbeige4.3-3B as one of the models that will come out of that distillation pipeline. It also says the 33B model is a cost-effective option to deploy directly.
The team compares the model with recent open models of about 30 billion parameters, including Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all run at their highest thinking setting with the same inference settings and harness. The comparison covers coding agent tests such as Terminal-Bench 2.1, Terminal-Bench 4.0, SWE-Atlas-QnA, NL2Repo and DeepSWE, and office and tool-use tests including Automation Bench, Toolathlon-Verified, EnterpriseOps-Gym, GDPval, Claw-Eval, DeepPlanning and Tau3-Bench. The team says Nanbeige4.3-33B leads on most of them and outperforms the baseline in most of the 35 occupations covered by JobBench. These results are self-reported and have not been independently verified.
The model supports a context window of up to 524,288 tokens, or 512K. Its chat template has an enable_thinking switch for reasoning mode and a preserve_thinking option that keeps earlier reasoning across turns, which the team recommends for multi-turn tool use, office tasks and code agents. Tool calls can use XML, the recommended format, or JSON. For agentic tasks the team suggests a temperature of 1.0 and up to 65,536 new tokens per response.
The model runs with Transformers, SGLang or vLLM, and the repository includes serving examples using a Qwen3-style reasoning parser and tool-call parser. The team says it is open to working with partners on agents for specific industries, including joint work on data and training.
The release adds another mid-sized option to the growing set of Chinese open-weight models aimed at coding agents rather than chat alone.
SEE ALSO: StepFun Releases Open Weights for Step 5 Preview · Kimi K3 Officially Open-Sourced