Kanzhun’s Nanbeige team has released Nanbeige4.3-33B on Hugging Face under the Apache 2.0 licence, Pandaily reported on 26 October. Kanzhun operates BOSS Zhipin, a Chinese online recruitment platform. The team presents the model as a way to run coding agents and office tasks directly, and as a larger model whose capabilities it plans to transfer into a smaller one.
Key points
- Nanbeige4.3-33B has 33 billion parameters and a context limit of 524,288 tokens.
- Nanbeige reports scores of 80.9 on Terminal-Bench 2.1 and 65.5 on DeepSWE in its own comparison.
- The Apache 2.0 release includes model weights that can be used commercially; the team plans a distilled 3B version.
Inside Nanbeige4.3-33B’s 524,288-token window
Nanbeige4.3-33B is a dense model: its parameters participate in processing each token, rather than being divided among specialist groups selected for different tokens. Its configuration specifies 52 layers, a hidden size of 7,168 and grouped-query attention with 56 query heads and eight key-value heads, MarkTechPost reported. That attention arrangement lets several query heads share a smaller set of key-value heads, reducing the information the system must retain while it works through a sequence.
The model accepts a context of up to 524,288 tokens, commonly described as 512K. Context is the material available to a model during a request: more room can accommodate earlier instructions, code and exchanges with tools alongside the current task. The chat template offers an enable_thinking setting for reasoning and a preserve_thinking option that retains earlier reasoning across turns. Nanbeige recommends the latter for multi-turn tool use, office work and coding agents, according to Pandaily.
Working through a long document could keep earlier passages within the stated context limit while later questions are put to the model. The preserve_thinking option could also retain reasoning from previous turns if it were enabled, matching the team’s stated approach to office tasks that continue across several exchanges.
Tool calls can be formatted as XML or JSON, with XML recommended by the team. For builders, those formats govern how the model asks another piece of software to act, while the reasoning settings govern what carries into the next exchange. The repository provides examples for Transformers, SGLang and vLLM; its server examples use four-way tensor parallelism, splitting the model’s work across four processors.
Terminal-Bench 2.1 and DeepSWE scores
Nanbeige tested its model against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B at each model’s highest thinking setting, using the same inference settings and test harness, Pandaily reported. The comparisons span coding-agent tasks, tool use, office work and reasoning. Pandaily described the results as self-reported and not independently verified.
In Nanbeige’s coding results, its model scores 80.9 on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, and 65.5 on DeepSWE against 42.5. The advantage does not extend to every coding test: Nanbeige’s table puts its SWE-Bench Pro score at 60.3, below Qwen3.8-27B’s 61.7. These are scores from the team’s specified benchmark runs, rather than measurements of coding work outside those tests.
Nanbeige’s model beats Qwen3.8-27B on seven of eight agentic-coding evaluations in the model-card comparison, Startup Fortune reported. Across the wider comparison with three models, it leads 17 of 20 benchmark rows and loses three, according to MarkTechPost. The counts cover different sets of tests; neither is a measure of every task an agent might encounter.
Competitor scores also require care. Nanbeige’s comparison table gives Gemma4-31B a GPQA Diamond score of 91.3, while Google lists 84.3 for that model, MarkTechPost reported. Differing test harnesses explain some gaps between the tables, and the affected rows should be treated as provisional pending independent runs.
Nanbeige’s planned 3B successor
Nanbeige calls the released model the teacher for a planned family: it intends to distil agent skills learned by the 33B version into smaller models, beginning with Nanbeige4.3-3B. Distillation means training a smaller model to reproduce useful behaviour from a larger one. The released weights give developers a model to run or adapt now, while the 3B version remains a plan described by the team.
The team has also released Nanbeige4.3-33B-Base for custom post-training. That gives builders a starting point distinct from the model configured for agent tasks, while the Apache 2.0 licence permits commercial use of the released weights. Nanbeige says it welcomes collaborators on sector-focused agents and on shared efforts involving datasets and model learning.
