Nanbeige4.3-33B is a 33-billion-parameter open language model released in October 2026 by the Nanbeige team at Kanzhun, the company behind the BOSS Zhipin recruitment app. It is published on Hugging Face under the Apache 2.0 license, supports a 512K-token context (524,288 tokens), and is built for agentic coding and what the team calls professional cowork tasks.
A teacher model first
The release strategy is unusual. The team says it built a high-capacity teacher agent first, and plans to distill its agent skills into much smaller models. Nanbeige4.3-33B is that flagship teacher; a far smaller student, Nanbeige4.3-3B, is named as next.
What the model card reports
The team compared it with Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, all at their highest thinking effort with the same settings and harness. These are self-reported results.
- Terminal-Bench 2.1: 80.9, against 73.0 for Qwen and 47.2 for Gemma.
- DeepSWE: 65.5, against 42.5 for Qwen, with the other two under 5.
- YC-bench, a simulated business: it ends with about $1.03 million, against roughly $641,000 for Qwen.
- JobBench: it beats the baseline in most of the 35 occupations tested.
Where it does not lead: Qwen3.8-27B edges it on SWE-Bench Pro (61.7 vs 60.3), Qwen leads GDPval (86.4 vs 84.8), and Gemma4-31B leads GPQA Diamond (91.3 vs 89.5).
How to run it
The model runs with Transformers, SGLang or vLLM. One tip from the quickstart: keep preserve_thinking off for normal chat, and turn it on for multi-turn tool use, office tasks and code-agent workflows. The team recommends temperature 1.0 for agentic tasks and 0.6 for reasoning and chat.
Our take
Self-reported numbers are a starting point, not a verdict. Test it on your own agent tasks before swapping models, and watch the planned small student: if a much smaller model keeps part of this skill, it could run agents on far cheaper hardware. For background, see our courses on agentic design patterns and on building agentic systems.
