Nanbeige's open 33B model beats Qwen on agentic coding

Nanbeige4.3-33B comes from Nanbeige LLM Lab, the AI team at Chinese job site BOSS Zhipin. The lab reports 65.5 on DeepSWE, well ahead of Qwen3.8-27B's 42.5. On Terminal-Bench 2.1 it reports 80.9 against Qwen's 73.0, Muse-Glimmer-30B's 61.8 and Gemma4-31B's 47.2. The weights went up on Hugging Face on 24 Oct under Apache 2.0.

This is a dense 33B model, not a mixture-of-experts, with a native 512K-token context. The lab says it is strongest at office-style agent work: 54.2 on Automation Bench against Qwen's 38.3, and 44.8 on JobBench against 34.5. It does not win everything. Qwen3.8-27B leads narrowly on SWE-Bench Pro (61.7 vs 60.3) and GDPval, and Gemma4-31B leads on GPQA Diamond (91.3 vs 89.5). All of these numbers come from the lab itself, and no independent evaluations have been published yet.

The bigger plan: the 33B model is a teacher. Nanbeige says it will distil it into smaller models, starting with Nanbeige4.3-3B. The lab is best known for 3B models, and that one has no release date yet.

Why it matters: a recruitment company now has an open model that competes with Qwen and Gemma on agent work, and you can run it commercially.