Nanbeige releases open-weights 33B model for agentic tasks

Nanbeige has released Nanbeige4.3-33B, a 33-billion-parameter open-weights model for agentic tasks, according to its Hugging Face model card1 and a November 5 Artificial Analysis post on X2.

Artificial Analysis said the dense model scores 20 on its Intelligence Index v5, the same as Qwen3.8 27B at its xhigh setting. It reported stronger terminal use but weaker knowledge, and said the model is small enough to run locally on a laptop. The independent results are published only in that post, which gives no run configuration beyond Qwen's xhigh setting.

Artificial Analysis describes Nanbeige as the foundation-model research team at BOSS Zhipin, the recruitment platform operated by Kanzhun Limited. The team's earlier Nanbeige4.2-3B used a Looped Transformer pretrained on 28 trillion tokens, according to its arXiv paper4. The model card calls the 33B model the "flagship teacher" of the Nanbeige4.3 series, to anchor distillation into smaller models, with Nanbeige4.3-3B given only as an example ("e.g.") of that pipeline. It also says the model delivers "state-of-the-art agentic performance among open models in the ~30B parameter class". That is a self-reported claim. Automatica reported on November 45 that Tencent released Hy3, a 295-billion-parameter open-weight mixture-of-experts model under Apache 2.0.

Artificial Analysis said the weights are Apache 2.0 and that there are no confirmed public first- or third-party APIs. The model takes text input and output only and has a 524,288-token context window, per the model card.

Terminal-Bench 4.0 shows the clearest gap between the two sets of figures. The model card reports 23.8 for Nanbeige against 6.3 for Qwen3.8-27B, and says all models ran at highest thinking effort with the same hyperparameters and harness. Artificial Analysis reports 30% against 6%. The sources do not explain the difference, and the model card's evaluation-results list marks the 23.8 figure with an asterisk without explanation.

On automation, Artificial Analysis scores Nanbeige at 53% against Qwen's 48% on AutomationBench-AA. The model card reports 54.2 against 38.3 on a benchmark it calls Automation Bench. The card also lists leads on Claw-Eval (78.3 against 75.2) and DeepPlanning (63.3 against 49.2). It shows Nanbeige behind Qwen on SWE-Bench Pro, 60.3 to 61.7. On GPQA Diamond, its 89.5 edges Qwen's 89.2 but trails Gemma4-31B's 91.3. Artificial Analysis also reports 16% for Nanbeige against 4% for Qwen on CritPt v2's Hard Subset.

Artificial Analysis found weaker results on coding and knowledge tests. Nanbeige scores 11% on FrontierCode 2.0 against Qwen's 15%, and 1,162 Elo on AA-Briefcase v1.1 against 1,391. It scores 7% on HLE-Diamond against 15%.

On AA-Omniscience v1.1, Nanbeige scores -4 against Qwen's -6. It declines 71% of questions and has a hallucination rate of 18% against Qwen's 29%. It answers 12% of questions correctly against 17%. Artificial Analysis said Nanbeige used 158,000 output tokens per Intelligence Index task, about 44% more than Qwen's 110,000.

The model card recommends temperature 1.0 and top_p 0.95 for agentic and tool-use tasks, and 0.6 and 0.95 for reasoning and chat. It says the model runs with Transformers, SGLang or vLLM.

Kanzhun also published a base checkpoint3. Its card reports 66.9 on MMLU-Pro against 59.3 for Gemma4-31B-Base.