Nanbeige4.3-33B: BOSS Zhipin's Apache 2.0 Model Beats Qwen3.8-27B on 18 of 20 Agent Benchmarks
Nanbeige LLM Lab, the AI research group inside BOSS Zhipin — China’s largest online recruiting platform — released Nanbeige4.3-33B on Hugging Face today under an Apache 2.0 license. The model targets agentic coding and professional task-completion workflows, and the lab’s benchmark table puts it ahead of Qwen3.8-27B on 18 of 20 tests. No independent verification has been published.
The model is a fine-tune of the Nanbeige4.3-33B-Base checkpoint, released in BF16. It ships with a 524,288-token context window and has reasoning enabled by default. A preserve_thinking flag carries reasoning state across tool calls, which the lab recommends for multi-step agent pipelines. The chat template supports both XML and JSON tool-call formats.
Benchmark Numbers
The widest leads are on terminal and software-agent tests:
| Benchmark | Nanbeige4.3-33B | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 80.9 | 73.0 | 61.8 | 47.2 |
| Terminal-Bench 4 | 23.8 | 6.3 | — | 0.0 |
| DeepSWE | 65.5 | 42.5 | — | — |
| Automation Bench | 54.2 | 38.3 | — | — |
| DeepPlanning | 63.3 | 49.2 | — | — |
| Toolathlon-Verified | 70.4 | 68.5 | — | — |
| JobBench | 44.8 | 34.5 | — | — |
| AIME 2026 | 95.2 | 94.5 | 94.7 | 89.2 |
| SciCode | 52.6 | — | — | — |
| AA-LCR long-context | 84.0 | — | — | — |
On YC-bench, which scores in simulated dollars earned, Nanbeige4.3-33B posts $1,025,437 against Qwen3.8-27B’s $641,290. The lab claims it beats the Qwen3.8-27B baseline across most of 35 occupational categories on JobBench.
The lead is not uniform. Qwen3.8-27B wins SWE-Bench Pro (61.7 vs 60.3) and GDPval (86.4 vs 84.8). Gemma4-31B leads GPQA Diamond at 91.3 vs Nanbeige’s 89.5.
Architecture and Deployment
The model runs on Transformers, SGLang, or vLLM. The lab recommends temperature 1.0 with up to 65,536 new tokens for agentic tasks, and temperature 0.6 with up to 262,144 new tokens for reasoning and long-form chat. Training data sources, training compute, and method are not disclosed in the model card.
Nanbeige4.3-33B is positioned as the teacher model for the Nanbeige4.3 series. The lab plans to distill it into a Nanbeige4.3-3B, following the now-standard pattern of using a large open-weight teacher to compress agentic capability into a deployable smaller model.
What It Is
BOSS Zhipin (NYSE: BZ) is a Beijing-based company with roughly 100 million registered users on its job-matching platform. Nanbeige LLM Lab is its AI research division, and this is the lab’s most capable public release. The name “Nanbeige” (南北阁, roughly “North-South Pavilion”) reflects the platform’s national reach inside China.
The model joins a growing class of mid-tier Apache 2.0 agent specialists — below the frontier on pure reasoning, but targeting the specific workloads where 30B-class compute budgets are practical: agentic pipelines, multi-step coding tasks, and automated professional workflows. The self-reported nature of every figure here warrants the usual caution until independent harnesses run the model.