# Nanbeige

> Source: https://aiwiki.ai/wiki/nanbeige
> Updated: 2026-11-17
> Fact-checked: 2026-11-17
> Categories: Chinese AI, Large Language Models, Open Source AI, Small Language Models
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Nanbeige." aiwiki.ai, 17 Nov 2026. https://aiwiki.ai/wiki/nanbeige
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Nanbeige** (Chinese: 南北阁, "South and North Pavilion") is a family of [open-weight](https://aiwiki.ai/wiki/open_weights) [large language models](https://aiwiki.ai/wiki/large_language_model) developed by the Nanbeige LLM Lab, the language-model research team of the Chinese online recruitment company BOSS Zhipin (Kanzhun Limited).[1][2] The current line began with Nanbeige4-3B in October 2025 and, as of November 2026, comprises three generations of 3B-class [small language models](https://aiwiki.ai/wiki/small_language_model) (Nanbeige4, Nanbeige4.1 and Nanbeige4.2) and one larger dense model, Nanbeige4.3-33B, released on October 25, 2026.[3][8] Every checkpoint in the current line is published on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license), which permits commercial use.[3]

The lab's stated focus has shifted over the line's history: the Nanbeige4 technical report concentrated on reasoning in a compact model, while the Nanbeige4.2 and Nanbeige4.3 releases describe themselves as agentic models for coding, tool use and office work.[5][6][7][4] Nanbeige4.3-33B scored 20.4 on version 5.0.1 of the [Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) Intelligence Index, against a median of 6 for open-weight models of similar size.[8]

## Developer

The organization card on Hugging Face describes Nanbeige as "a large language model (LLM) research team at BOSS Zhipin".[1] The technical reports for Nanbeige4-3B and Nanbeige4.2-3B list the authors' affiliation as "Nanbeige LLM Lab" and "Boss Zhipin", and the model cards give a contact address on the kanzhun.com domain.[5][7][4] Kanzhun Limited, which operates the BOSS Zhipin recruitment platform, is listed on Nasdaq (BZ) and the Hong Kong Stock Exchange (2076).[2] As of November 2026 the Hugging Face organization listed 19 members.[1]

Kanzhun has referred to the models in its financial results. Its fourth-quarter and full-year 2025 release, dated March 18, 2026, described Nanbeige as the company's "self-developed large language model" and said it had ranked first among Hugging Face's trending text models.[2] In the first-quarter 2026 release of May 20, 2026, founder and chief executive Jonathan Peng Zhao said the open-sourced Nanbeige 4.1-3B had been "ranked jointly No. 1 among sub-4B parameter models" by Artificial Analysis.[9] Both statements are the company's own.

The lab also maintains a GitHub organization, created in October 2023, that holds forks of the [vLLM](https://aiwiki.ai/wiki/vllm), [SGLang](https://aiwiki.ai/wiki/sglang), [llama.cpp](https://aiwiki.ai/wiki/llama_cpp) and [Ollama](https://aiwiki.ai/wiki/ollama) inference engines patched for its models, and an architecture research repository, Depth Hybrid Attention, published in October 2026.[10][11]

## History

### Earlier generations

The Nanbeige name predates the current line. Nanbeige-16B, a 16-billion-parameter model pretrained on 2.5 trillion tokens, was released with Base, Chat and 32K-context variants; third-party quantized copies of it appeared on Hugging Face in November 2023.[12] Its original model card required users to comply with both the Apache 2.0 license and a separate "License Agreement for Large Language Models Nanbeige", and asked anyone intending commercial use to apply by email for a revocable commercial license.[12] Quantized copies of 8B and 16B models from the Nanbeige1.5 and Nanbeige2 lines followed in 2024.[13] As of November 2026, the organization's public Hugging Face listing contains only fourth-generation and later models.[3]

The Nanbeige4 technical report names Nanbeige3.5-Pro, a model that has not been published on Hugging Face, as the teacher used when distilling Nanbeige4-3B and as a source of reference answers during reinforcement learning.[5][3]

### Fourth generation

| Model | Hub date | Parameters (safetensors total) | Context (config) | Type | Weights license |
| --- | --- | --- | --- | --- | --- |
| Nanbeige4-3B-Thinking-2510 | October 22, 2025 | same configuration as Nanbeige4-3B-Base | 65,536 | Reasoning | Apache 2.0[3][14] |
| Nanbeige4-3B-Base | October 31, 2025 | 3,933,637,120 | 32,768 | Base | Apache 2.0[3][15] |
| Nanbeige4-3B-Thinking-2511 | November 20, 2025 | 3,933,637,120 | 65,536 | Reasoning | Apache 2.0[3][16] |
| Nanbeige4.1-3B | February 10, 2026 | 3,933,637,120 | 262,144 | Reasoning and agentic | Apache 2.0[3][6] |
| Nanbeige4.2-3B-Base | July 21, 2026 | 4,169,800,704 | 131,072 | Base (looped) | Apache 2.0[3] |
| Nanbeige4.2-3B | July 21, 2026 | 4,169,800,704 | 262,144 | Agentic, hybrid thinking | Apache 2.0[3][7] |
| Nanbeige4.3-33B-Base | October 24, 2026 | 33,118,977,024 | 131,072 | Base | Apache 2.0[3][17] |
| Nanbeige4.3-33B | October 24, 2026 (announced October 25) | 33,118,977,024 | 524,288 | Agentic, hybrid thinking | Apache 2.0[3][4] |

"Hub date" is the creation date of the Hugging Face repository; public announcements can trail it by a day or more.[3][18] Nanbeige4.1-3B is dated February 11, 2026 by Artificial Analysis, and its technical report was posted to [arXiv](https://aiwiki.ai/wiki/arxiv) on February 13.[19][6] The Nanbeige4.2-3B report was posted on July 24, 2026.[7] Nanbeige4.3-33B was announced on the lab's X account; Artificial Analysis dates the release to October 25, 2026, and DoNews reported the announcement on October 26.[8][18]

The organization also published quantized checkpoints (FP8 and GPTQ Int8 versions of Nanbeige4.2-3B in August 2026; FP8, INT8 and INT4 versions of Nanbeige4.3-33B on October 31, 2026), a draft model called Nanbeige4.2-3B-DSpark for [speculative decoding](https://aiwiki.ai/wiki/speculative_decoding) (August 31, 2026), a search agent called ToolMind-Web-3B, fine-tuned from Nanbeige4-3B-Thinking-2511 (January 2026), and CoSineVerifier-Tool-4B, an answer verifier trained from Alibaba's Qwen3-4B-Instruct-2507 and released under the MIT License.[3][20][21][22]

## Architecture and training

### Nanbeige4-3B

Nanbeige4-3B uses a standard dense decoder configuration loaded as `LlamaForCausalLM`: 32 layers, a hidden size of 2,560, [grouped-query attention](https://aiwiki.ai/wiki/grouped_query_attention) with 4 key-value heads, and a 166,144-token vocabulary.[23] According to the technical report, the base model was pretrained on 23 trillion tokens of web text, books, code and papers, supplemented with synthetic question-answer, textbook and long [chain-of-thought](https://aiwiki.ai/wiki/chain_of_thought) data.[5] Training followed what the lab calls a Fine-Grained Warmup-Stable-Decay (FG-WSD) schedule: a 0.1T-token warmup, two stable stages of 12.4T and 6.5T tokens with progressively higher-quality data mixtures, and a 4T-token decay and long-context stage weighted toward math, code and synthetic reasoning data.[5][15]

Post-training used more than 30 million [supervised fine-tuning](https://aiwiki.ai/wiki/fine_tuning) samples, followed by [knowledge distillation](https://aiwiki.ai/wiki/knowledge_distillation) from Nanbeige3.5-Pro through a method the authors call Dual Preference Distillation, which combines token-level distillation with sequence-level preference optimization.[5] Three stages of on-policy [GRPO](https://aiwiki.ai/wiki/grpo) reinforcement learning then targeted STEM reasoning, coding and human-preference alignment.[5] The November 2025 update (Thinking-2511) was described as a further round of distillation and [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) over the October release.[16]

### Nanbeige4.1-3B

Nanbeige4.1-3B kept the Nanbeige4-3B-Base weights as its starting point and added further supervised fine-tuning and reinforcement learning.[6] The report describes point-wise and pair-wise reward modeling for alignment, complexity-aware rewards for code that score efficiency as well as correctness, and synthesized deep-search data with turn-level supervision.[6] The developers said the model could sustain up to 600 tool-call turns on a single problem; the model card gives the figure as "more than 500 rounds".[6][24] For [deep research](https://aiwiki.ai/wiki/deep_research) tasks the card recommends running it in the MiroFlow framework from MiroMind's MiroThinker project.[24]

### Nanbeige4.2-3B

Nanbeige4.2-3B introduced a new base model, pretrained from scratch on 28 trillion tokens with a looped transformer that passes hidden states through the same layer stack twice, adding depth without adding parameters.[7] Its configuration has 22 layers run for 2 loops, a hidden size of 3,072 and 8 key-value heads, under a custom `NanbeigeForCausalLM` class.[25] The lab counts about 3B non-embedding parameters and 4B in total.[7] Supervised fine-tuning ran as a three-stage curriculum at 64K, 128K and 256K context, followed by reinforcement learning over both thinking and non-thinking responses, length-controlled reasoning RL, and agentic RL that combines outcome and process rewards.[7] The model card says the accompanying modeling code also contains features planned for a future Nanbeige4.5 (LoopSplit, mHC with depth attention, and concatenated n-gram embeddings) and that Nanbeige4.5 was in training for release later in 2026.[26]

### Nanbeige4.3-33B

Nanbeige4.3-33B is a dense model loaded as `LlamaForCausalLM`, with 52 layers, a hidden size of 7,168, 56 attention heads with 8 key-value heads, a 166,912-token vocabulary, and a configured context of 524,288 tokens.[27] The base model's configuration lists 131,072 positions.[17] No technical report had been published as of November 17, 2026, and the model card does not describe the training data or recipe.[4] The lab presents it as a "flagship teacher model" for the Nanbeige4.3 series, intended to distill agentic capabilities into smaller models, beginning with a planned Nanbeige4.3-3B.[4] That smaller model had not appeared on Hugging Face as of November 17, 2026.[3] The chat template switches reasoning on or off (`enable_thinking`) and can keep earlier reasoning in multi-turn conversations (`preserve_thinking`).[4]

## Licensing

Every repository in the fourth-generation line carries the `apache-2.0` license tag in its model-card metadata, including the base models and quantized variants.[3] The repositories do not include a separate LICENSE file, and the model cards do not add use restrictions beyond a general disclaimer about harmful outputs.[3][26] Under the Apache 2.0 terms, commercial use, modification and redistribution are permitted. This is a change from Nanbeige-16B, whose card required an application for a commercial license.[12] The CoSineVerifier-Tool-4B verifier is the one exception in the current organization: it is tagged MIT.[3] The training-data releases ToolMind, ToolMind-Web-QA and SSR-RCoT-16K are tagged Apache 2.0, but the pretraining corpora are not published.[30][5]

## Benchmark results

### Developer-reported

The following figures come from Nanbeige's own model cards and reports, with evaluation harnesses chosen by the developer.

For Nanbeige4.3-33B, the lab compared the model with Alibaba's [Qwen3.8-27B](https://aiwiki.ai/wiki/qwen3_8), Meta's [Muse-Glimmer-30B](https://aiwiki.ai/wiki/muse_glimmer) and Google's [Gemma4-31B](https://aiwiki.ai/wiki/gemma_4), each at its highest thinking effort, and reported the best score on 17 of 20 benchmarks.[4][18]

| Benchmark (Nanbeige4.3-33B card) | Nanbeige4.3-33B | Qwen3.8-27B | Muse-Glimmer-30B | Gemma4-31B |
| --- | --- | --- | --- | --- |
| [Terminal-Bench](https://aiwiki.ai/wiki/terminal_bench) 2.1 | 80.9 | 73.0 | 61.8 | 47.2 |
| [SWE-bench](https://aiwiki.ai/wiki/swe_bench) Pro | 60.3 | 61.7 | 51.2 | 36.9 |
| [GDPval](https://aiwiki.ai/wiki/gdpval) | 84.8 | 86.4 | 75.8 | 71.4 |
| [GPQA](https://aiwiki.ai/wiki/gpqa) Diamond | 89.5 | 89.2 | 83.5 | 91.3 |
| [AIME](https://aiwiki.ai/wiki/aime) 2026 | 95.2 | 94.5 | 94.7 | 89.2 |
| JobBench | 44.8 | 34.5 | 16.6 | 9.0 |

Source: Nanbeige4.3-33B model card.[4] The three benchmarks where it did not lead were SWE-bench Pro, GDPval and GPQA Diamond.[4]

For Nanbeige4.2-3B, the lab reported 63.6 on SWE-bench Verified, 44.1 on Terminal-Bench 2.0 and 87.4 on GPQA Diamond, ahead of Qwen3.5-9B and Gemma4-12B on most tasks in its table.[26] Gemma4-12B scored higher in the same table on SciCode, [IFBench](https://aiwiki.ai/wiki/ifbench) and Recruit-Bench, an in-house benchmark of hiring and job-seeking scenarios.[26] Nanbeige4.1-3B was reported at 76.9 on [LiveCodeBench](https://aiwiki.ai/wiki/livecodebench)-V6 and 73.2 on [Arena-Hard](https://aiwiki.ai/wiki/arena_hard)-v2, and Nanbeige4-3B-Thinking-2510 at 87.5 on AIME 2024 and 77.2 on GPQA, each compared with [Qwen3](https://aiwiki.ai/wiki/qwen_3) models of similar or larger size.[24][14] The base-model comparisons in the Nanbeige4 report were made by fine-tuning each base model on the same public datasets rather than by few-shot testing.[15]

### Independent evaluations

Artificial Analysis scored Nanbeige4.3-33B at 20.4 on its Intelligence Index v5.0.1, a composite of nine evaluations, placing it well above the median of 6 for open-weight models in its size class.[8] Alibaba's Qwen3.8 27B, at its Xhigh reasoning setting, scored 20.1 on the same index version.[28] Artificial Analysis also recorded that Nanbeige4.3-33B generated 160 million output tokens while running the index, against a class median of 77.5 million, and described it as "very verbose".[8] Artificial Analysis listed Nanbeige4.1-3B with a February 11, 2026 release date, but its page showed no v5.0.1 index score as of November 2026.[19]

In an August 24, 2026 study of small models running on an iPhone 17 Pro, Artificial Analysis found Nanbeige4.2-3B tied for the highest average score (63) on its five-evaluation mobile benchmark set at a 16K context limit, level with [Liquid AI](https://aiwiki.ai/wiki/liquid_ai)'s LFM2.5-2.6B.[29] The same study measured Nanbeige4.2-3B as slower and larger in memory on the phone than LFM2.5-2.6B, scored it 33% on a non-hallucination measure against 79% for the Liquid model, and found its score fell to 18 when each answer was limited to one minute of generation time.[29] The phone speed and memory measurements in that study were developed and run by Liquid AI; the intelligence scores were administered by Artificial Analysis.[29]

## Availability

All checkpoints are downloadable from the Nanbeige organization on Hugging Face, and none is gated.[3] As of November 2026, Artificial Analysis listed no publicly available API for Nanbeige4.3-33B and reported no speed or cost figures for it.[8]

Nanbeige4.3-33B uses a Llama-compatible configuration, and its card gives serving examples for Transformers, SGLang and vLLM with four-GPU [tensor parallelism](https://aiwiki.ai/wiki/tensor_parallelism).[4] Nanbeige4.2-3B's looped architecture is not in the upstream inference engines, so its card points to the lab's own branches of SGLang, vLLM, llama.cpp and Ollama, names a published Ollama model, and warns that LM Studio's bundled llama.cpp server does not yet support it.[26][10] The card for Nanbeige4.2-3B also suggests using it as a local personal assistant with the [OpenClaw](https://aiwiki.ai/wiki/openclaw) agent framework.[26] Within days of the Nanbeige4.3-33B release, third parties had posted [GGUF](https://aiwiki.ai/wiki/gguf) quantizations and modified variants to Hugging Face, including versions with safety tuning removed.[13]

## Reception

Coverage of Nanbeige has come from Chinese and English-language technology media and from Kanzhun's own investor communications. DoNews reported the Nanbeige4.3-33B release as BOSS Zhipin's large-model lab open-sourcing a 33.1B model, noting that the performance claims were the developer's own and listing the benchmarks where it trailed rivals, such as SWE-bench Pro, GDPval and GPQA Diamond, and occupations in JobBench, such as healthcare management and customer service, where it trailed.[18] The Nanbeige4-3B-Thinking-2511 card cites placements on third-party leaderboards, including 25th overall on the Berkeley Function Calling Leaderboard ([BFCL](https://aiwiki.ai/wiki/bfcl)) and 11th on [WritingBench](https://aiwiki.ai/wiki/writingbench) at the time of release.[16]

Several caveats apply to the published results. Most scores come from harnesses chosen by the developer, the office and co-work results for Nanbeige4.2-3B use an in-house scaffold, and Recruit-Bench was built by the lab itself.[26][7] Artificial Analysis's figures support a high index score for Nanbeige4.3-33B relative to its size, but they also show heavy token use, and its phone study found that Nanbeige4.2-3B was far more likely to guess than to decline on questions it could not answer.[8][29]

## See also

- [Qwen](https://aiwiki.ai/wiki/qwen)
- [MiniCPM](https://aiwiki.ai/wiki/minicpm)
- [Small language model](https://aiwiki.ai/wiki/small_language_model)
- [List of open-weight models](https://aiwiki.ai/wiki/list_of_open_weight_models)
- [Open source AI](https://aiwiki.ai/wiki/open_source_ai)

## References

1. Nanbeige LLM Lab. "Nanbeige LLM Lab" (organization card). Hugging Face, retrieved November 17, 2026. https://huggingface.co/Nanbeige
2. KANZHUN LIMITED. "KANZHUN LIMITED Announces Fourth Quarter and Full Year 2025 Financial Results". GlobeNewswire / Kanzhun investor relations, March 18, 2026. https://kanzhun.gcs-web.com/news-releases/news-release-details/kanzhun-limited-announces-fourth-quarter-and-full-year-2025
3. Hugging Face. "Models by Nanbeige" (Hub API listing with repository dates and license tags). Hugging Face, retrieved November 17, 2026. https://huggingface.co/api/models?author=Nanbeige&full=true&limit=100
4. Nanbeige Team. "Nanbeige4.3-33B" (model card). Hugging Face, October 2026. https://huggingface.co/Nanbeige/Nanbeige4.3-33B
5. Yang, Chen; et al. "Nanbeige4-3B Technical Report: Exploring the Frontier of Small Language Models". arXiv:2512.06266, December 6, 2025. https://arxiv.org/abs/2512.06266
6. Yang, Chen; et al. "Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts". arXiv:2602.13367, February 13, 2026. https://arxiv.org/abs/2602.13367
7. Nanbeige LLM Lab; et al. "Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Model". arXiv:2607.22083, July 24, 2026 (revised July 27, 2026). https://arxiv.org/abs/2607.22083
8. Artificial Analysis. "Nanbeige4.3-33B: Intelligence, Performance & Price Analysis". Artificial Analysis, retrieved November 17, 2026. https://artificialanalysis.ai/models/nanbeige4-3-33b
9. KANZHUN LIMITED. "KANZHUN LIMITED Announces First Quarter 2026 Financial Results". GlobeNewswire / Kanzhun investor relations, May 20, 2026. https://kanzhun.gcs-web.com/news-releases/news-release-details/kanzhun-limited-announces-first-quarter-2026-financial-results
10. Nanbeige LLM Lab. "Nanbeige" (GitHub organization and repositories). GitHub, retrieved November 17, 2026. https://github.com/Nanbeige
11. Nanbeige LLM Lab. "Depth Hybrid Attention" (README). GitHub, October 2026. https://github.com/Nanbeige/Depth-Hybrid-Attention
12. TheBloke. "Nanbeige 16B Chat - GGUF" (includes the original Nanbeige-16B-Chat model card). Hugging Face, November 2023. https://huggingface.co/TheBloke/Nanbeige-16B-Chat-GGUF
13. Hugging Face. "Models matching 'nanbeige'" (search listing including third-party quantizations). Hugging Face, retrieved November 17, 2026. https://huggingface.co/models?search=nanbeige
14. Nanbeige LLM Lab. "Nanbeige4-3B-Thinking-2510" (model card). Hugging Face, October 2025. https://huggingface.co/Nanbeige/Nanbeige4-3B-Thinking-2510
15. Nanbeige LLM Lab. "Nanbeige4-3B-Base" (model card). Hugging Face, October 2025. https://huggingface.co/Nanbeige/Nanbeige4-3B-Base
16. Nanbeige LLM Lab. "Nanbeige4-3B-Thinking-2511" (model card). Hugging Face, November 2025. https://huggingface.co/Nanbeige/Nanbeige4-3B-Thinking-2511
17. Nanbeige Team. "Nanbeige4.3-33B-Base" (model card and config.json). Hugging Face, October 2026. https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base
18. Yang Liang. "BOSS Zhipin's Nanbeige open-sources Nanbeige4.3-33B" (in Chinese: "BOSS 直聘旗下 Nanbeige 开源 Nanbeige4.3-33B"). DoNews, October 26, 2026. https://www.donews.com/news/detail/1/6753392.html
19. Artificial Analysis. "Nanbeige4.1-3B" (model page). Artificial Analysis, retrieved November 17, 2026. https://artificialanalysis.ai/models/nanbeige4-1-3b
20. Nanbeige LLM Lab. "Nanbeige4.2-3B-DSpark" (model card). Hugging Face, August 2026. https://huggingface.co/Nanbeige/Nanbeige4.2-3B-DSpark
21. Nanbeige LLM Lab. "ToolMind-Web-3B" (model card). Hugging Face, January 2026. https://huggingface.co/Nanbeige/ToolMind-Web-3B
22. Feng, Ruixiang; et al. "CoSineVerifier: Tool-Augmented Answer Verification for Computation-Oriented Scientific Questions". arXiv:2512.01224, December 1, 2025. https://arxiv.org/abs/2512.01224
23. Nanbeige LLM Lab. "config.json", Nanbeige4-3B-Base. Hugging Face. https://huggingface.co/Nanbeige/Nanbeige4-3B-Base/blob/main/config.json
24. Nanbeige LLM Lab. "Nanbeige4.1-3B" (model card). Hugging Face, February 2026. https://huggingface.co/Nanbeige/Nanbeige4.1-3B
25. Nanbeige LLM Lab. "config.json", Nanbeige4.2-3B. Hugging Face. https://huggingface.co/Nanbeige/Nanbeige4.2-3B/blob/main/config.json
26. Nanbeige LLM Lab. "Nanbeige4.2-3B" (model card). Hugging Face, July 2026 (updated September 2026). https://huggingface.co/Nanbeige/Nanbeige4.2-3B
27. Nanbeige Team. "config.json", Nanbeige4.3-33B. Hugging Face. https://huggingface.co/Nanbeige/Nanbeige4.3-33B/blob/main/config.json
28. Artificial Analysis. "Qwen3.8 27B" (model page). Artificial Analysis, retrieved November 17, 2026. https://artificialanalysis.ai/models/qwen3-8-27b
29. Artificial Analysis. "Intelligence at pocket scale: Benchmarking small models and mobile phones". Artificial Analysis, August 24, 2026. https://artificialanalysis.ai/articles/mobile-phone-intelligence-inference
30. Hugging Face. "Datasets by Nanbeige" (Hub API listing). Hugging Face, retrieved November 17, 2026. https://huggingface.co/api/datasets?author=Nanbeige&full=true

