BOSS Zhipin open-sources the 3B agent model Nanbeige 4.2, outperforming larger models in tests.

According to monitoring by Beating, Boss Zhipin’s Nanbeige Lab has open-sourced Nanbeige4.2-3B. With only about 3 billion parameters, it can simultaneously modify code, process office documents, invoke complex tools, and operate terminals. Official tests show that it outperforms larger models Qwen3.5-9B and Gemma4-12B across multiple general agent and code agent benchmarks. On SWE-Bench Verified, it scores 63.6, compared to 53.1 for Qwen3.5-9B and 44.2 for Gemma4-12B. Nanbeige4.2-3B repeats the computation of the same layer group twice, trading increased computational cost for enhanced capability—enabling smaller models to handle more complex tasks without increasing parameter count, though requiring more computing power and time for inference. Nanbeige Lab aims to reduce the continuous invocation cost of agents using smaller models. Nanbeige4.2-3B has released its model weights and supports Transformers, SGLang, vLLM, llama.cpp, and Ollama.