Nanbeige4.3-33B matches Qwen3.8 27B at 20 on the Artificial Analysis Intelligence Index v5

Nanbeige4.3-33B matches Qwen3.8 27B at 20 on the Artificial Analysis Intelligence Index v5

Nanbeige, the foundation model research team at BOSS Zhipin, the Chinese recruitment platform operated by Kanzhun Limited, has released Nanbeige4.3-33B. The dense 33B reasoning model is available as open weights under the Apache 2.0 license.

The text-only model accepts text input and produces text output. It has a 524K-token context window and is small enough to run locally on a laptop. No confirmed public first-party or third-party APIs are available.

Key results

• Competitive performance at around 30B parameters: Nanbeige4.3-33B and Qwen3.8 27B (xhigh) both score 20 on the Artificial Analysis Intelligence Index v5. Muse Glimmer (high, 30B) scores 6.
• Relative strengths in terminal use and physics reasoning: Nanbeige4.3-33B scores 30% on Terminal-Bench 4.0 and 16% on CritPt v2's Hard Subset, compared with 6% and 4%, respectively, for Qwen3.8 27B.
• Different evaluation strengths despite the same overall score: Nanbeige scores 53% on AutomationBench-AA, compared with 48% for Qwen3.8 27B. It trails Qwen3.8 27B on AA-Briefcase v1.1, with 1162 versus 1391 Elo; FrontierCode 2.0, with 11% versus 15%; and HLE-Diamond, with 7% versus 15%.
• High abstention and lower hallucination on Omniscience: Nanbeige scores -4 on AA-Omniscience v1.1, compared with -6 for Qwen3.8 27B. Its hallucination rate is 18%, versus 29%. However, it answers fewer questions correctly, 12% versus 17%, and declines 71% of questions.
• More output tokens per task: Nanbeige uses 158K output tokens per Intelligence Index task, compared with 110K for Qwen3.8 27B, approximately 44% more, while both models score 20.

Model details

• Type: Open-weights reasoning model
• License: Apache 2.0
• Size: 33B total parameters, dense
• Modalities: Text input and output only
• Context window: 524K tokens
• API availability: No confirmed public first-party or third-party APIs

Intelligence Index performance and parameter count

Nanbeige4.3-33B scores 20 on the Intelligence Index with 33B total parameters. Qwen3.8 27B also scores 20 with 27B parameters. Both are dense models, establishing a new Pareto-optimal point for Intelligence Index score versus total parameters.

DeepSeek V4.1 Flash 1017 (max) scores 33 with 552B total parameters. Its mixture-of-experts architecture activates 16B parameters per token.

Selective answering on AA-Omniscience

Nanbeige4.3-33B's AA-Omniscience v1.1 result reflects a more selective answering pattern. Its hallucination rate is 18%, compared with 29% for Qwen3.8 27B, while its attempt rate is 29%, compared with 43%.

Qwen3.8 27B answers more questions correctly overall, at 17% versus 12% for Nanbeige4.3-33B. Nanbeige declines 71% of questions, indicating that its overall result includes a high rate of abstention.

Output-token usage

Nanbeige4.3-33B uses 158K output tokens per Intelligence Index task, compared with 110K for Qwen3.8 27B. This is approximately 44% more output despite the identical overall score of 20.

The figure is a weighted average of answer and reasoning tokens across the nine v5 evaluations, including turns involving agentic tool use.