Nanbeige releases open weights for its coding agent model
Nanbeige has released its coding model and a matching base checkpoint under Apache 2.0. The laboratory behind the Chinese recruitment platform BOSS Zhipin is targeting software tasks that involve files, tools and tests. The release also provides a teacher model for planned transfers of capabilities into smaller systems, while leaving developers able to perform their own post-training.
Artificial Intelligence··Night
Two checkpoints under Apache
Nanbeige, the language-model laboratory within Kanzhun, the company operating China’s BOSS Zhipin recruitment platform, has released Nanbeige4.3-33B with a matching base model. Both are distributed through Hugging Face, a platform for sharing machine-learning models. Actu.ai and AI Bacon report that the weights carry the Apache 2.0 licence, allowing commercial use, modification and redistribution under its terms.[1], [2]
The release separates a post-trained system from the checkpoint preceding that additional training. AI Bacon describes the finished model as supporting thinking and tool calling, while the base version is intended for developers conducting their own post-training. Access to the weight files gives developers the model’s learned numerical parameters; it does not establish that all of the material used to train it has also been published.[1], [2]
Coding through files and tools
The model targets coding agents: software that can work through a programming assignment in several steps. The coding workflow includes reading project files, running terminal commands and executing tests before revising a proposed change. These are the kinds of multi-step tasks the release is intended to support, extending beyond a single suggested completion or an answer about a fragment of code. The laboratory identifies English and Chinese as the model’s supported languages. Its dense architecture uses all model parameters for each token, the units into which text is divided for processing. The available model family therefore includes a system ready for further application work and a base checkpoint for additional training, with both versions distributed under the same licence.[1], [2]
A teacher for smaller models
Nanbeige also presents the larger model as a teacher for its smaller systems. In distillation, a smaller model is trained to reproduce aspects of a larger model’s responses. The laboratory’s plan names Nanbeige4.3-3B as an intended recipient. This is a proposed transfer of capabilities within the model family, rather than a report that the smaller system already reproduces the larger model’s abilities.[1], [2]
AI Bacon dates the Hugging Face release to October 24; Actu.ai’s October 26 report places it around that date. The laboratory’s earlier releases included Nanbeige4.2-3B. AI Bacon describes the larger checkpoint as arriving before the smaller model it is intended to teach, with the smaller member planned as a recipient of its coding-agent capabilities.[1], [2]
