Recently, SuperCLUE released the latest OnDevice mobile phone-side large model evaluation ranking. In this competition focusing on the local operation capabilities of smartphones, vivo's BlueLM3.5Nano3B model performed impressively, securing the top position in the overall ranking with a score of 89.86.

Looking at the specific evaluation results, this 3B parameter small model not only topped the list in the edge-side field, but its score is also approaching that of major cloud-based large models such as Google's Gemini3.6Flash, Doubao Seed2.1pro, and Qwen3.8Max. In the subsequent rankings of the edge-side list, Qwen3.59B and Qwen3.54B scored 87.82 and 85.16 respectively, placing second and third. Other participants include GLM4.6V Flash, Gemma4E4B it, and Gemma4E2B it.

image.png

With the accelerating popularization of edge-side AI, the advantages of local operation mode are gradually becoming apparent. The core value of edge-side large models lies in their ability to perform image and text reasoning directly on the phone without requiring a constant internet connection. All user input content remains on the device, providing a more secure and private experience.

Although constrained by the physical limitations of smartphone hardware, the parameter size of edge-side models cannot be infinitely scaled like cloud-based models, and they still have limitations when handling ultra-complex tasks. However, vivo's achievement of an excellent score with a 3B parameter model proves that small-volume edge-side models can also achieve a high technical ceiling. This suggests that future smartphone AI experiences will no longer rely excessively on the network, allowing smooth operation even in offline environments. Of course, benchmark scores only reflect laboratory performance, and actual daily usage effects still depend on the deep system optimization by major smartphone manufacturers.