Nanbeige
Nanbeige (Chinese: 南北阁, "South and North Pavilion") is a family of open-weight large language models 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 (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 under the Apache License 2.0, 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 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, SGLang, llama.cpp and 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 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 (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 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 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 samples, followed by 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 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 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 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, Meta's Muse-Glimmer-30B and Google's Gemma4-31B, 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 2.1 | 80.9 | 73.0 | 61.8 | 47.2 |
| SWE-bench Pro | 60.3 | 61.7 | 51.2 | 36.9 |
| GDPval | 84.8 | 86.4 | 75.8 | 71.4 |
| GPQA Diamond | 89.5 | 89.2 | 83.5 | 91.3 |
| 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 and Recruit-Bench, an in-house benchmark of hiring and job-seeking scenarios.[26] Nanbeige4.1-3B was reported at 76.9 on LiveCodeBench-V6 and 73.2 on Arena-Hard-v2, and Nanbeige4-3B-Thinking-2510 at 87.5 on AIME 2024 and 77.2 on GPQA, each compared with Qwen3 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'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.[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 agent framework.[26] Within days of the Nanbeige4.3-33B release, third parties had posted 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) and 11th on 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]
Continue reading
- Small language modelA small language model (SLM) is a compact large language model, typically under about 10 billion parameters, built to run efficiently on a single consumer GPU, a laptop, or a smartphone rather than in a data…Listed in See also
- QwenQwen is a family of large language models and multimodal models developed by the Qwen Team at Alibaba Cloud.Listed in See also
- MiniCPMMiniCPM is a family of compact, openly licensed language models published by OpenBMB, the shared open-source brand of Tsinghua University's natural language processing lab (THUNLP) and the Beijing company…Listed in See also
- List of open-weight modelsThis is a list of open-weight models: artificial intelligence models whose trained parameters (weights) the developer has published for download, so that the model can be run, studied and in most cases…Listed in See also
- Open-source AIOpen-source AI refers to artificial intelligence systems whose recipients have broad rights to use, study, modify, and share the system, together with access to the materials needed to exercise those rights.Listed in See also
- Hugging FaceHugging Face is an artificial intelligence company and collaboration platform for machine learning.Mentioned 11 times above
More to read
Topics:Chinese AILarge Language ModelsOpen Source AISmall Language Models
References
- ^1 ^2 ^3Nanbeige LLM Lab. "Nanbeige LLM Lab" (organization card). Hugging Face, retrieved November 17, 2026. huggingface.co/Nanbeige
- ^1 ^2 ^3KANZHUN LIMITED. "KANZHUN LIMITED Announces Fourth Quarter and Full Year 2025 Financial Results". GlobeNewswire / Kanzhun investor relations, March 18, 2026. kanzhun.gcs-web.com/...-quarter-and-full-year-2025
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19Hugging Face. "Models by Nanbeige" (Hub API listing with repository dates and license tags). Hugging Face, retrieved November 17, 2026. huggingface.co/...models
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10Nanbeige Team. "Nanbeige4.3-33B" (model card). Hugging Face, October 2026. huggingface.co/...Nanbeige4.3-33B
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Yang, Chen; et al. "Nanbeige4-3B Technical Report: Exploring the Frontier of Small Language Models". arXiv:2512.06266, December 6, 2025. arxiv.org/...2512.06266
- ^1 ^2 ^3 ^4 ^5 ^6Yang, Chen; et al. "Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts". arXiv:2602.13367, February 13, 2026. arxiv.org/...2602.13367
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Nanbeige LLM Lab; et al. "Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Model". arXiv:2607.22083, July 24, 2026 (revised July 27, 2026). arxiv.org/...2607.22083
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Artificial Analysis. "Nanbeige4.3-33B: Intelligence, Performance & Price Analysis". Artificial Analysis, retrieved November 17, 2026. artificialanalysis.ai/...nanbeige4-3-33b
- ^KANZHUN LIMITED. "KANZHUN LIMITED Announces First Quarter 2026 Financial Results". GlobeNewswire / Kanzhun investor relations, May 20, 2026. kanzhun.gcs-web.com/...rter-2026-financial-results
- ^1 ^2Nanbeige LLM Lab. "Nanbeige" (GitHub organization and repositories). GitHub, retrieved November 17, 2026. github.com/Nanbeige
- ^Nanbeige LLM Lab. "Depth Hybrid Attention" (README). GitHub, October 2026. github.com/...Depth-Hybrid-Attention
- ^1 ^2 ^3TheBloke. "Nanbeige 16B Chat - GGUF" (includes the original Nanbeige-16B-Chat model card). Hugging Face, November 2023. huggingface.co/...Nanbeige-16B-Chat-GGUF
- ^1 ^2Hugging Face. "Models matching 'nanbeige'" (search listing including third-party quantizations). Hugging Face, retrieved November 17, 2026. huggingface.co/models
- ^1 ^2Nanbeige LLM Lab. "Nanbeige4-3B-Thinking-2510" (model card). Hugging Face, October 2025. huggingface.co/...Nanbeige4-3B-Thinking-2510
- ^1 ^2 ^3Nanbeige LLM Lab. "Nanbeige4-3B-Base" (model card). Hugging Face, October 2025. huggingface.co/...Nanbeige4-3B-Base
- ^1 ^2 ^3Nanbeige LLM Lab. "Nanbeige4-3B-Thinking-2511" (model card). Hugging Face, November 2025. huggingface.co/...Nanbeige4-3B-Thinking-2511
- ^1 ^2Nanbeige Team. "Nanbeige4.3-33B-Base" (model card and config.json). Hugging Face, October 2026. huggingface.co/...Nanbeige4.3-33B-Base
- ^1 ^2 ^3 ^4Yang 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. donews.com/...6753392
- ^1 ^2Artificial Analysis. "Nanbeige4.1-3B" (model page). Artificial Analysis, retrieved November 17, 2026. artificialanalysis.ai/...nanbeige4-1-3b
- ^Nanbeige LLM Lab. "Nanbeige4.2-3B-DSpark" (model card). Hugging Face, August 2026. huggingface.co/...Nanbeige4.2-3B-DSpark
- ^Nanbeige LLM Lab. "ToolMind-Web-3B" (model card). Hugging Face, January 2026. huggingface.co/...ToolMind-Web-3B
- ^Feng, Ruixiang; et al. "CoSineVerifier: Tool-Augmented Answer Verification for Computation-Oriented Scientific Questions". arXiv:2512.01224, December 1, 2025. arxiv.org/...2512.01224
- ^Nanbeige LLM Lab. "config.json", Nanbeige4-3B-Base. Hugging Face. huggingface.co/...config.json
- ^1 ^2 ^3Nanbeige LLM Lab. "Nanbeige4.1-3B" (model card). Hugging Face, February 2026. huggingface.co/...Nanbeige4.1-3B
- ^Nanbeige LLM Lab. "config.json", Nanbeige4.2-3B. Hugging Face. huggingface.co/...config.json
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Nanbeige LLM Lab. "Nanbeige4.2-3B" (model card). Hugging Face, July 2026 (updated September 2026). huggingface.co/...Nanbeige4.2-3B
- ^Nanbeige Team. "config.json", Nanbeige4.3-33B. Hugging Face. huggingface.co/...config.json
- ^Artificial Analysis. "Qwen3.8 27B" (model page). Artificial Analysis, retrieved November 17, 2026. artificialanalysis.ai/...qwen3-8-27b
- ^1 ^2 ^3 ^4Artificial Analysis. "Intelligence at pocket scale: Benchmarking small models and mobile phones". Artificial Analysis, August 24, 2026. artificialanalysis.ai/...ne-intelligence-inference
- ^Hugging Face. "Datasets by Nanbeige" (Hub API listing). Hugging Face, retrieved November 17, 2026. huggingface.co/...datasets
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 2,737 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Checked by an independent verifier against the Hugging Face repositories and configuration files, the arXiv technical reports, Kanzhun's earnings releases and Artificial Analysis; 3 material and 9 minor fixes applied before publishing.
Cite this page: AI Wiki. "Nanbeige." aiwiki.ai, updated 17 Nov 2026, fact-checked 17 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/nanbeige


