BOSS Zhipin has open-sourced the 3B Agent model, officially tested to surpass Qwen3.5 9B and Gemma4 12B.

According to Foresight Beating monitoring, Nanbeige Lab under BOSS Zhipin has open-sourced Nanbeige4.2-3B. It has only about 3 billion parameters, but it can simultaneously modify code, process office documents, call complex tools, and operate terminals.

Official tests show that in multiple general Agent and code Agent evaluations, it outperforms the larger-parameter models Qwen3.5-9B and Gemma4-12B. On SWE-Bench Verified, it scored 63.6 points, Qwen3.5-9B scored 53.1 points, and Gemma4-12B scored 44.2 points.

Nanbeige4.2-3B will make the same group of model layers calculate twice, trading more computation for greater capability. This way, without increasing parameters, smaller models can handle more complex tasks, but inference requires more computing power and time.

Nanbeige Lab hopes to reduce the continuous invocation cost of Agents with small models. Nanbeige4.2-3B has released model weights, supporting Transformers, SGLang, vLLM, llama.cpp, and Ollama.