Nanbeige releases 33B open-source agent model on Hugging Face
Nanbeige4.3-33B ships under Apache 2.0 with a 524,288-token context window, and Nanbeige claims wins over Qwen3.8-27B on 17 of 20 tests.
Nanbeige, a lab operated by the Chinese recruiting platform BOSS Zhipin, released the open-source model Nanbeige4.3-33B on Hugging Face on October 24, according to aitimes.com. The model carries 33 billion parameters and is published under the Apache 2.0 license, which permits commercial and research use as well as modification and redistribution of the weights.
Nanbeige4.3-33B handles both English and Chinese and uses a dense architecture rather than a mixture-of-experts design. It supports a context window of up to 524,288 tokens. The model card lists run instructions for Transformers, SGLang and vLLM, and notes that the BF16 weights alone require about 66GB of memory.
Nanbeige says the model scored highest on 17 of 20 evaluation items against Alibaba's Qwen3.8-27B, Meta's Muse-Glimmer-30B and Google's Gemma4-31B. On Terminal-Bench 2.1 it scored 80.9 against 73.0 for Qwen3.8-27B, on DeepSWE 65.5 against 42.5, on Automation Bench 54.2 against 38.3 and on JobBench 44.8 against 34.5. In the YC-Bench startup simulation, the company reported simulated revenue of $1,025,437 versus $641,290 for Qwen3.8-27B.
The comparison is not a clean sweep. Nanbeige4.3-33B trailed Qwen3.8-27B on SWE-bench Pro, with 60.3 against 61.7, and on GDPval, with 84.8 against 86.4.
Nanbeige said it plans to use Nanbeige4.3-33B as a teacher model for distillation and is preparing a smaller follow-up, Nanbeige4.3-3B.
Quick answers
What license is Nanbeige4.3-33B released under?
The Apache 2.0 license, which allows commercial and research use as well as modification and redistribution of the weights.
How large is Nanbeige4.3-33B's context window?
It supports up to 524,288 tokens.
How much memory do the BF16 weights need?
About 66GB for the BF16 weights alone.