Intel / article

Nanbeige Has Released Nanbeige4 3 33b Scoring 20 On The Intelligence Index

Source: artificialanalysis.ai
Author: Artificial Analysis articles
Date: Published Nov 5, 2026

Context

Nanbeige4.3-33B, a new open-weights from Nanbeige, the foundation-model research team at Chinese recruitment platform BOSS Zhipin, ties Qwen's Qwen3.8 27B at 20 on the Artificial Analysis Intelligence Index v5, but the two reach that number in different ways (Artificial Analysis). The index is an independent composite of benchmark results that was rebuilt in late October so that most of it rests on private test sets, with agentic coding and science tasks added (earlier coverage of v5). In those tests Nanbeige scored 30% on Terminal-Bench 4.0, which has a model work through tasks at a command line, against 6% for Qwen, and 16% against 4% on a hard physics-reasoning subset of CritPt v2. It also edged Qwen on AutomationBench-AA, 53% to 48%. Qwen was ahead on FrontierCode 2.0 (15% to 11%), on the HLE-Diamond knowledge and reasoning exam (15% to 7%) and on AA-Briefcase.

The model also behaves more cautiously on factual questions. On AA-Omniscience v1.1 it declined 71% of questions and answered only 12% correctly, against 17% for Qwen, while its rate was 18% against Qwen's 29%. It gets there with more text: Artificial Analysis counted 158,000 output tokens per index task, including reasoning, against 110,000 for Qwen, about 44% more, which means more compute and waiting time for the same overall score.

Nanbeige4.3-33B is a dense model with 33 billion parameters, released under the Apache 2.0 license, with text input and output only and a of 524K tokens. Artificial Analysis says it is small enough to run locally on a laptop, and that it found no confirmed public API from Nanbeige or a third party, so for now running it yourself is the way to use it. For comparison, DeepSeek's V4.1 Flash scores 33 but has 552 billion total parameters, though its design activates only 16 billion per token.

Key takeaways · AI-distilled

Nanbeige4.3-33B comes from the foundation model team at recruitment platform BOSS Zhipin; it is a dense 33B reasoning model under Apache 2.0, text only, with a 524K-token context and no confirmed public API.
Despite tying Qwen3.8 27B, Nanbeige scores 30% vs 6% on Terminal-Bench 4.0 and 16% vs 4% on CritPt v2's Hard Subset, but trails on FrontierCode 2.0 (11% vs 15%) and HLE-Diamond (7% vs 15%).
On AA-Omniscience, Nanbeige declines 71% of questions, giving an 18% hallucination rate vs 29% for Qwen3.8 27B, though it answers fewer correctly overall (12% vs 17%).
Artificial Analysis measured 158K output tokens per Intelligence Index task for Nanbeige vs 110K for Qwen3.8 27B, about 44% more for the same score of 20.

Terms in this piece · Glossary

reasoning model — A model trained to think — generating extended internal reasoning before answering — trading time and tokens for accuracy on hard problems.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.

Read the source artificialanalysis.ai