**[BOSS Zhipin Open-Sources 3B Agent Model, Outperforms Qwen3.5 9B and Gemma4 12B in Official Tests]** According to monitoring by Dongcha Beating, BOSS Zhipin's Nanbeige Lab has open-sourced the Nanbeige4.2-3B model. Despite having only about 3 billion parameters, it is capable of simultaneously modifying code, processing office documents, invoking complex tools, and operating terminals. Official tests show that it surpasses larger parameter models like Qwen3.5-9B and Gemma4-12B in multiple general Agent and code Agent evaluations. On the SWE-Bench Verified benchmark, it scored 63.6 points, compared to 53.1 points for Qwen3.5-9B and 44.2 points for Gemma4-12B. Nanbeige4.2-3B achieves its enhanced capabilities by having the same model layer compute twice, trading additional computation for greater performance. This approach allows the smaller model to handle more complex tasks without increasing the number of parameters, though it requires more computational power and time for inference. Nanbeige Lab aims to use smaller models to reduce the continuous invocation costs of Agents. The Nanbeige4.2-3B model weights have been made publicly available and support frameworks such as Transformers, SGLang, vLLM, llama.cpp, and Ollama. [Original Link]