Nanbeige releases open-weight 33B model scoring 20 on Artificial Analysis index
The dense 33B Nanbeige4.3-33B is released under Apache 2.0 and scores 20 on the Artificial Analysis Intelligence Index, level with Qwen3.8 27B.
Nanbeige, the foundation model research team at BOSS Zhipin, the Chinese recruitment platform operated by Kanzhun Limited, has released Nanbeige4.3-33B as open weights under the Apache 2.0 license. Artificial Analysis puts it at 20 on its Intelligence Index v5, level with Qwen3.8 27B (xhigh) and well ahead of Muse Glimmer (high, 30B) at 6. The model is dense, meaning all 33B parameters are used for every token. Artificial Analysis sets it beside DeepSeek V4.1 Flash 1017 (max), which scores 33 but has 552B total parameters and uses a mixture-of-experts design that activates only 16B per token. Artificial Analysis says Nanbeige4.3-33B sets a new Pareto-optimal point for Intelligence Index against total parameters.
The Intelligence Index is a composite of nine v5 evaluations. Artificial Analysis says terminal use is a relative strength. The model scores 30% on Terminal-Bench 4.0, which tests agentic coding and work in a terminal, against 6% for Qwen3.8 27B. It also scores 16% on the Hard Subset of CritPt v2, a physics reasoning test, against 4% for Qwen3.8 27B. Nanbeige's model card reports 23.8 on Terminal-Bench 4.0 against 6.3 for Qwen3.8-27B, evaluated with mini-swe-agent v2.4.6 at the highest thinking effort. Those model-card figures differ from the Artificial Analysis numbers, so the two sets should not be merged.
The model is not uniformly ahead of Qwen3.8 27B. Artificial Analysis reports that Nanbeige4.3-33B beats it on AutomationBench-AA, 53% to 48%. It trails on AA-Briefcase v1.1 (1162 to 1391 Elo), FrontierCode 2.0 (11% to 15%) and HLE-Diamond (7% to 15%).
On AA-Omniscience v1.1, Nanbeige4.3-33B scores -4 against -6 for Qwen3.8 27B. The methodology on Artificial Analysis's model page explains that the net score rewards correct answers, penalizes hallucinations and does not penalize refusals. The model's hallucination rate is 18% against 29%, but it answers fewer questions correctly, 12% against 17%, and declines 71% of questions. Its attempt rate is 29% against 43%. Artificial Analysis says the lower hallucination rate reflects selective answering.
Token use is also high. Nanbeige4.3-33B used 158K output tokens per Intelligence Index task, against 110K for Qwen3.8 27B, about 44% more, even though both score 20. Artificial Analysis says the figure is a weighted average of answer and reasoning tokens across the nine evaluations and includes agentic tool-use turns. The model page lists the model as very verbose, with 160M output tokens from the Intelligence Index against a median of 77.7M for comparable open-weights models.
No public APIs are confirmed yet. Artificial Analysis says there are no confirmed public first- or third-party APIs, and its model page shows no cost per task. The weights are on Hugging Face, where the model card gives Transformers, vLLM and SGLang instructions and lists a context length of 524,288 tokens, which it calls 512K and Artificial Analysis gives as 524K.
The model card positions the model as the flagship teacher for a planned distillation pipeline and describes it as delivering state-of-the-art agentic performance among open models in the roughly 30B parameter class. The Artificial Analysis model page gives the release date as October 25, 2026, while the article is dated November 5, 2026.
What changed
Nanbeige4.3-33B is a new dense open-weight 33B model released under Apache 2.0 that scores 20 on the Artificial Analysis Intelligence Index.
Type: New capability · Applies to: Nanbeige4.3-33B
Why it matters
Teams that want a small open model for local or self-hosted use now have another option at this score level, with a clear trade-off between terminal-use strength and factual recall. The higher token use per task also affects inference cost.
What you can do
Benchmark it on your own terminal and knowledge tasks before adopting it, since the strengths and weaknesses are task-specific.
Original source
Further reporting
Where we saw it
Artificial Analysis
AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works