BOSS Zhipin's AI Lab Publishes a Coding Model Anyone Can Download and Reuse
Nanbeige LLM Lab has published the weights of its 33-billion-parameter coding model, Nanbeige4.3-33B, on Hugging Face under the Apache 2.0 license, which allows commercial use and modification. The lab is the model team at Kanzhun, the Chinese company behind the BOSS Zhipin hiring platform, and it created the repository on Oct. 24, 2026.
The lab describes the model as the flagship teacher of its Nanbeige4.3 series: a large model built first so that its abilities can be transferred into smaller ones, a process called distillation. A 3-billion-parameter version is planned. The lab says the model is meant for complex coding and professional office tasks.
The repository's own metadata records the model at 33 billion parameters and about 66 gigabytes of files to download, and the model card gives a context length of up to 524,288 tokens. A window that large is what lets a model read an entire code repository or a book in a single prompt. The card lists the settings the lab recommends for agent work and for ordinary chat, and the model runs under the common open-source serving tools.
The card also carries a table of scores on 20 benchmarks of coding, office work and reasoning, placing Nanbeige4.3-33B beside other recently released models of about the same size. The lab says it leads on most of them. Those scores are the lab's own: it ran every model itself, using the same test harness and settings for each. No independent evaluation of the model has been published.
Nanbeige4.3-33B is a further-trained version of the lab's own Nanbeige4.3-33B-Base, and it handles English and Chinese. No paper came with the release, and the citation the lab suggests points to the Hugging Face page itself.
