BOSS Zhipin's Nanbeige Lab Open-Sources Nanbeige4.3-33B Agent Model Under Apache 2.0
The 33B dense model with 512K context claims Terminal-Bench 2.1 score of 80.9 and is positioned as a teacher for smaller distillations.
Nanbeige, the AI lab of Chinese recruitment platform BOSS Zhipin (Kanzhun), released Nanbeige4.3-33B, a dense 33.1-billion-parameter model with a Llama-style architecture and a 524,288-token context window, licensed under Apache 2.0 for unrestricted commercial use. The lab positions it as a teacher model to be distilled into smaller variants such as a planned Nanbeige4.3-3B.
The model card reports Terminal-Bench 2.1 at 80.9 versus 73.0 for Qwen3.8-27B, and DeepSWE at 65.5 versus 42.5, while trailing slightly on SWE-Bench Pro and GPQA Diamond. All figures are self-reported using the same harness across compared models. Community testing indicates the model runs quantised on consumer-class GPUs such as the RTX 4090 and Apple Silicon, making it a realistic candidate for single-GPU self-hosting of agentic document workflows.