Nanbeige4.3-33B ties Qwen3.8 27B on Intelligence Index
BOSS Zhipin's Nanbeige team has released an open-weights 33B dense reasoning model that scores 20 on Artificial Analysis's Intelligence Index v5, a new best score for its total parameter count.
Nanbeige4.3-33B matches Qwen3.8 27B (xhigh) at 20 on the Intelligence Index v5, Artificial Analysis reported. Nanbeige is the foundation-model research team at BOSS Zhipin, the Chinese recruitment platform run by Kanzhun Limited.
The model is a dense 33B reasoning model released under Apache 2.0. It handles text input and output only, has a 524K-token context window and is small enough to run locally on a laptop. Artificial Analysis says no public first- or third-party API has been confirmed.
The two models reach the same score in different ways. Nanbeige leads Qwen3.8 27B on Terminal-Bench 4.0 (30% vs. 6%), the CritPt v2 Hard Subset (16% vs. 4%) and AutomationBench-AA (53% vs. 48%). It trails on AA-Briefcase v1.1 (1162 vs. 1391 Elo), FrontierCode 2.0 (11% vs. 15%) and HLE-Diamond (7% vs. 15%).
On AA-Omniscience v1.1, Nanbeige answers selectively. It declines 71% of questions and has a hallucination rate of 18%, against 29% for Qwen. That gives it a score of -4 to Qwen's -6, although it answers fewer questions correctly (12% vs. 17%).
The model is also less token-efficient. It used 158K output tokens per Index task, compared with 110K for Qwen3.8 27B, about 44% more.
Artificial Analysis says the result sets a new Pareto-optimal point for Intelligence Index score against total parameters. Mixture-of-experts models still score higher at low active-parameter counts. DeepSeek V4.1 Flash 1017 (max), for example, scores 33 while activating 16B of its 552B parameters per token.