BOSS Zhipin releases 3B agent model with a reused transformer stack

BOSS Zhipin's Nanbeige LLM Lab released the Nanbeige4.2-3B agent model on July 24th, using a Looped Transformer architecture that reuses the same transformer layers twice to increase effective depth without adding parameters. The 3-billion-parameter model, trained on 28 trillion tokens, scored 63.6 on SWE-bench Verified and outperformed larger models like Qwen3.5-9B and Gemma4-12B in agent evaluations, according to the lab's own benchmarks.

Nanbeige LLM Lab https://huggingface.co/Nanbeige/Nanbeige4.2-3B , the AI research group inside Chinese recruiting platform BOSS Zhipin, released a compact agent model on July 24th that uses the same transformer layers twice to increase its effective depth without adding parameters. The release puts BOSS Zhipin https://ir.zhipin.com/about-us/company-profile , founded by Peng Zhao https://ir.zhipin.com/corporate-governance/management , in an unusual position among applied internet platforms. BOSS Zhipin built Nanbeige for hiring workflows, then expanded the research program into general models intended to run coding agents, office tools and personal assistants. Zhao previously led rival recruitment platform Zhaopin and holds a law degree from Peking University. The 24-author research team, led on the paper by Chen Yang, published the Nanbeige4.2-3B technical report https://arxiv.org/abs/2607.22083 alongside Apache-2.0 model weights https://huggingface.co/Nanbeige/Nanbeige4.2-3B . Nanbeige4.2-3B has 4 billion total parameters, including 3 billion outside its embedding layers, and supports English and Chinese with a context window of up to 262,144 tokens. The release spread beyond Nanbeige's own channels on July 27th when Tanishq Mathew Abraham, Ph.D. @iScienceLuvr https://x.com/iscienceluvr/status/2081657822345826600?s=46 posted the model and paper on X. Abraham is not affiliated with Nanbeige. Smaller weights, repeated computation Nanbeige trained the base model from scratch on 28 trillion tokens, increasing the share of mathematics, code and synthetic question-answering data in the mixture. The corpus also included a smaller amount of agent-trajectory data, according to the paper. Its main architectural bet is the Looped Transformer. Hidden states travel through the model's layer stack and then pass through those same layers a second time. That gives the model greater effective computational depth while keeping its parameter count fixed. The parameter savings still carry a compute cost. Nanbeige found that the two-pass setup retained about 75% of the token efficiency of a standard transformer. Additional passes produced smaller gains, slower training and less stable optimization. The lab also tested sharing the key-value cache between passes, which cut cache requirements in half but reduced performance, and chose the full non-sharing configuration for the released model. Post-training combined supervised fine-tuning with several reinforcement-learning stages. Nanbeige assembled software-engineering, tool-use and office-task trajectories from real and synthetic environments, then filtered them using execution results and rubric-based assessments. Its reinforcement-learning process targeted hallucinations, repetitive output, instruction errors, reasoning length and failures during multi-step tool use. The model card provides instructions for running Nanbeige4.2-3B through Transformers, vLLM and SGLang. Nanbeige also published paths for GGUF and int4 quantization, plus modified branches of llama.cpp and Ollama for local inference. That packaging matters for developers evaluating small models because nominal parameter count alone does not make a model easy to deploy. Strong vendor benchmarks, with caveats Nanbeige reports that the model scored 63.6 on SWE-bench Verified, 46.9 on SWE-bench Pro and 44.1 on Terminal-Bench 2.0. It also reported 87.4 on GPQA-Diamond and 52.2 on Claw-Eval. In the lab's comparisons, Nanbeige4.2-3B beat larger Qwen3.5-9B and Gemma4-12B models across every listed general-agent and code-agent evaluation. It also led five of six reasoning tests in the comparison table. Gemma4-12B remained ahead on two alignment evaluations, including BOSS Zhipin's internal Recruit-Bench. Those figures are Nanbeige's own evaluations. Some office and coworking tests used an in-house scaffold, while Recruit-Bench was designed internally around employer and job-seeker tasks. Nanbeige says it used established agent frameworks for the major coding benchmarks, including OpenHands for SWE-bench Verified and SWE-agent for SWE-bench Pro. The training design also exposes where the compact model remains constrained. During agentic reinforcement learning, Nanbeige favored relatively easy tasks with shorter trajectories and higher pass rates because those examples produced more stable training and larger gains than difficult, long-horizon tasks. The model card separately warns that outputs can contain inaccurate, biased or harmful material. BOSS Zhipin turns an internal model into a broader platform bet BOSS Zhipin began deploying Nanbeige in 2024 for search, recommendations, AI-assisted communication and interviews. The recruiting platform combines the proprietary model with third-party foundation models, according to Kanzhun's 2025 annual filing https://ir.zhipin.com/static-files/3580737f-e15c-414d-ad5b-432850a867f8 . The model program sits under a technical organization overseen by Tao Zhang https://ir.zhipin.com/board-member/tao-zhang/ , BOSS Zhipin's chief technology officer /article/rob-witoff-joins-coinbase-as-cto since its inception. Zhang previously worked at IBM group companies, Renren and Baidu and is responsible for research, development and IT infrastructure. BOSS Zhipin has the scale to fund the work internally. Kanzhun reported 60.9 million average monthly active users in the first quarter of 2026 and 7.1 million paid enterprise customers during the 12 months ended March 31st. First-quarter revenue reached RMB2.07 billion, or $299.9 million, while research and development expenses were RMB423.8 million, according to its May 20th results https://www.sec.gov/Archives/edgar/data/1842827/000110465926064167/tm2614941d2 ex99-1.htm . Kanzhun has also told investors that Nanbeige requires continuing capital and personnel spending and may never generate enough direct revenue to cover those costs. Open-sourcing Nanbeige4.2-3B gives BOSS Zhipin another return on that investment: external adoption, feedback and technical credibility for a model program that began inside a recruiting app. For agent developers, the practical test is whether the reported performance survives outside Nanbeige's evaluation setup. A capable 3-billion-parameter model can reduce memory requirements and make private, low-latency deployments more accessible, even when its looped architecture demands extra computation per token.