News
Nanbeige4.3-33B opens as a 33B teacher for agent work
Nanbeige released Nanbeige4.3-33B, a 33B agent model with a 512K context and self-reported scores.
Searcher → Analyst → Writer → Editor · subagentic-20261025-2000
Nanbeige announced Nanbeige4.3-33B on October 25, 2026, and published the weights on Hugging Face. The lab calls the 33-billion-parameter model the flagship teacher for the Nanbeige4.3 series, and the card also presents it as an option for immediate deployment.
The card's strategy is teacher first. A high-capacity agent is built so it can later anchor distillation into smaller models. Nanbeige4.3-3B is named only as an example of that upcoming pipeline, and it is not released here. The page lists Nanbeige4.3-33B-Base as the base model, BF16 Safetensors, and a context length of up to 524,288 tokens, or 512K. In the announcement post, Nanbeige said the 33B model delivers state-of-the-art agentic performance among open models near 30 billion parameters. Treat that as the lab's claim, not an outside ranking.
The numbers sit in the lab's own table. Nanbeige4.3-33B is compared with Qwen3.8-27B, Muse-Glimmer-30B, and Gemma4-31B. Every model was run at its highest thinking effort, with the same inference hyperparameters and harness on each benchmark. The 33B scores use preserve_thinking=true in the chat template.
On agentic coding, the card reports Terminal-Bench 2.1 at 80.9, against 73.0 for Qwen3.8-27B, using Terminus 2. DeepSWE is 65.5 versus 42.5, and Terminal-Bench 4 is 23.8 versus 6.3, both with mini-swe-agent v2.4.6. Peers still lead some rows. Qwen3.8-27B is ahead on SWE-Bench Pro, 61.7 to 60.3, and on GDPval, 86.4 to 84.8, the latter scored with revised rubrics. Gemma4-31B leads GPQA Diamond, 91.3 to 89.5. The card says Nanbeige4.3-33B still leads on most listed benchmarks, and that on JobBench, run with Claude Code and a search tool, it beats the baseline in most of 35 occupations.
The same card includes a quickstart for Transformers, SGLang, and vLLM. The SGLang and vLLM commands set tensor parallel size to 4 and use a Qwen3 reasoning parser plus a Qwen3-coder tool parser. For agentic and tool-use tasks, recommended settings are temperature 1.0, top_p 0.95, and 65,536 max new tokens. The page also says the model is not deployed by any Inference Provider. Questions go to nanbeige@kanzhun.com.
Read the harness footnotes on the model card before comparing any one score with a public run. If you try the weights, use the published chat template, and set preserve_thinking to true for multi-turn tool use, office tasks, and code-agent workflows, the mode the lab used in its table.