Nanbeige4.3-33B ties Qwen3.8 27B at 20 on Artificial Analysis — and beats it badly on Terminal-Bench 4.0

Nanbeige, the foundation-model team at recruitment platform BOSS Zhipin (Kanzhun), has an open-weight dense reasoner that Artificial Analysis now puts level with Alibaba's Qwen3.8 27B — with a very different shape underneath the same composite score.

Nanbeige4.3-33B shipped on Hugging Face around October 25 under Apache 2.0: 33B dense parameters, text-only, 524K context, no confirmed public API. On November 5, Artificial Analysis published an independent run on Intelligence Index v5: both Nanbeige4.3-33B and Qwen3.8 27B (xhigh) score 20. That ties them well above Muse Glimmer 30B (high) at 6 in AA's same-size open-weight comparison.

Same index, different strengths

AA's breakdown is the interesting part:

• Terminal-Bench 4.0: Nanbeige 30% vs Qwen3.8 27B 6% (AA harness).
• CritPt v2 (Hard): 16% vs 4%.
• AutomationBench-AA: 53% vs 48%.
• FrontierCode 2.0: Nanbeige 11%, Qwen 15% — coding quality goes the other way.
• AA-Omniscience: Nanbeige answers fewer questions correctly (12% vs 17%) but declines 71% of them, cutting the hallucination rate to 18% vs 29%. Net Omniscience: −4 vs −6.

Nanbeige is also wordier on the index suite: about 158K output tokens per Intelligence Index task vs 110K for Qwen3.8 27B (~44% more).

The lab's own model card is more aggressive on agent coding. With Terminus 2 it reports 80.9% on Terminal-Bench 2.1 against 73.0 for Qwen3.8-27B, and 23.8% vs 6.3% on Terminal-Bench 4.0 under mini-swe-agent — directionally the same Terminal-Bench story AA saw, with vendor numbers that still need treating as claims. (We covered a separate Terminal-Bench 2.0→2.1 repair audit today; AA's Index uses Terminal-Bench 4.0.)

What it is for

Nanbeige frames 4.3-33B as a teacher for a forthcoming smaller distilled line (they name Nanbeige4.3-3B). The card recommends preserve_thinking=true for multi-turn tool use and code agents, Transformers/SGLang/vLLM recipes, and XML tool-call formatting. Cost on AA is blank: you pay for hardware.

A 33B Apache-2.0 dense model that matches a strong Qwen peer on the composite while dominating Terminal-Bench is worth a look for local agent work. The knowledge vs terminal-use trade-off is the number to remember, not the shared "20".