Clarity for AI buildersSponsored by AI Routing
availability

Nanbeige releases open-weight 33B model scoring 20 on Artificial Analysis Intelligence Index

Nanbeige releases open-weight 33B model scoring 20 on Artificial Analysis Intelligence Index

Nanbeige has released Nanbeige4.3-33B, an open-weight reasoning model that scores 20 on the Artificial Analysis Intelligence Index v5, matching Qwen3.8 27B (xhigh), according to Artificial Analysis. Its evaluation shows stronger terminal-use and physics-reasoning results than Qwen, but weaker performance on several knowledge and coding measures and higher output-token use.

Nanbeige is the foundation-model research team at BOSS Zhipin, the Chinese recruitment platform operated by Kanzhun Limited. Artificial Analysis describes the new model as a dense, 33-billion-parameter system released under the Apache 2.0 license, with text-only input and output and a 524K-token context window. The source reports no confirmed public first- or third-party API availability.

Different strengths behind the same score

The shared Intelligence Index score masks substantial differences across individual evaluations. According to Artificial Analysis, Nanbeige4.3-33B scores 30% on Terminal-Bench 4.0, compared with 6% for Qwen3.8 27B, and 16% on the Hard Subset of CritPt v2, compared with 4%. It also leads on AutomationBench-AA, at 53% versus 48%.

Qwen performs better on other measures. Nanbeige records 1162 Elo on AA-Briefcase v1.1 against Qwen's 1391, 11% on FrontierCode 2.0 against 15%, and 7% on HLE-Diamond against 15%. These results suggest that the equal overall score should not be treated as equivalent performance across tasks.

Lower hallucination rate, more abstention

Artificial Analysis reports that Nanbeige scores -4 on AA-Omniscience v1.1, compared with -6 for Qwen. Nanbeige's reported hallucination rate is lower, at 18% versus 29%, but it also attempts fewer questions: 29%, compared with Qwen's 43%.

Nanbeige declines 71% of questions and answers 12% correctly overall. Qwen answers 17% correctly. The lower hallucination rate therefore comes alongside more selective answering, rather than a higher overall rate of correct responses.

Higher output-token use

Nanbeige uses 158K output tokens per Intelligence Index task, compared with 110K for Qwen—about 44% more despite the same overall score. Artificial Analysis says this is a weighted average across the index's nine v5 evaluations, including answer tokens, reasoning tokens and agentic tool-use turns.

The results give users a task-dependent comparison: Nanbeige leads on the reported terminal-use, physics-reasoning and automation tests, while Qwen leads on the cited coding, knowledge and AA-Briefcase measures. The supplied evidence does not establish comparative latency or operating costs, so the token-use difference alone does not settle deployment efficiency.