Nanbeige releases Nanbeige4.3-33B, an open ~30B agentic model

Nanbeige announced Nanbeige4.3-33B on X on October 25, with the weights and a base model (Nanbeige4.3-33B-Base) on Hugging Face. The lab says it built a strong "teacher" model first, to push the ceiling of small agents, and that the 33B will teach upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.

The model is released under Apache 2.0, supports English and Chinese, and has a context length of up to 512K tokens. The model card gives vLLM and SGLang launch commands and recommends keeping the model's reasoning across turns (preserve_thinking) for tool use and code-agent work.

Claims from the announcement, not independently measured:

• State-of-the-art agentic performance among open models at about 30B parameters.
• Strong results on agentic coding and cowork benchmarks. In the lab's own table against Qwen3.8-27B, Muse-Glimmer-30B and Gemma4-31B, it scores 80.9 on Terminal-Bench 2.1 (Qwen3.8-27B: 73.0) and 65.5 on DeepSWE (42.5), and trails Qwen3.8-27B on SWE-Bench Pro (60.3 vs 61.7) and GDPval (84.8 vs 86.4).
• Handles real-world tasks with cost-efficient inference when used in mainstream agent harnesses.

The post had about 1,300 likes on X by October 25. One user relaying it said the benchmark scores look clearly higher than Qwen 3.8 27B's, which is their reading of the lab's chart.