SenseTime SenseNova V6.5 Multimodal Large Model Ranked No.1 in China in 2025

Claiming the title of No.1 in China, SenseTime SenseNova V6.5 triumphed in the year-end showdown of multimodal large models in 2025. 

Recently, SuperCLUE, an authoritative large model evaluation benchmark, released the December Report on Chinese Multimodal Vision-Language Model Evaluation Benchmark. SenseTime SenseNova V6.5 (SenseNova V6.5 Pro) secured first place in China with an overall score of 75.35 and received the highest score in the visual reasoning dimension. 

SenseTime SenseNova V6.5 Leads the Chinese Camp, Highlighting Its Global Competitiveness 

The evaluation covered three core dimensions—basic cognition, visual reasoning and visual application—along with 20 sub-tasks. 

The evaluation results showed that the latest version of SenseTime SenseNova V6.5 Pro ranked first among Chinese models with a score of 75.35, outperforming competitors such as Doubao-seed-1.6-vision, ERNIE-5.0 and Qwen3. Furthermore, Chinese models like SenseTime SenseNova are rapidly narrowing the gap with leading international models, marking a significant leap in their global competitiveness. 

In the dimension-based capability comparison, the report pointed out that Chinese models, including SenseTime SenseNova V6.5, have approached the average level of top-tier models in basic cognition, with some even on par with Gemini 3.0 Pro. In the visual reasoning dimension, SenseTime SenseNova V6.5 is the only Chinese model to surpass the average level of leading models, placing it at the forefront of the industry, while other Chinese models still face a notable gap in this area. 

Specifically, SenseTime SenseNova V6.5 achieved first place in China in six sub-tasks – object description, text recognition, environment identification, logical reasoning, code design, and autonomous driving scenarios – and achieved the highest domestic score of 79.17 in the scientific reasoning task. 

SenseTime SenseNova V6.5 also delivered outstanding performance in sub-tasks such as facial recognition, 3D object recognition, mathematical reasoning, industrial application capability, medical image analysis, graphical user interface understanding, and content moderation. It firmly ranks among the first echelon of Chinese models, demonstrating comprehensive multimodal cognitive, reasoning, and application capabilities. 

Adhering to Foundamental Technology Innovation, Driving Model Iteration with User Value

The breakthrough in the multimodal field is a powerful testament to SenseTime’s dedication to the "multimodal general intelligence" technology strategy and its commitment to breaking industrial development bottlenecks through foundational innovation. 

From innovative model architectures and the removal of data barriers to the reform of training paradigms, SenseTime’s SenseNova large models have continuously expanded the capability boundaries of multimodal large models. 

This year, SenseTime also released and open-sourced the NEO multimodal architecture, which redefines the fusion mechanism of vision and language, enabling inherent multimodal capabilities. It has achieved optimal performance comparable to peer multimodal models with only 1/10 the data volume typically required. By integrating multimodal data and adopting the groundbreaking Cross-View Prediction training paradigm, which goes beyond Next Token Prediction, SenseTime SenseNova has surpassed GPT-5, the latest Gemini-3 Pro, and Cambrian-S in spatial intelligence performance. In addition, based on joint innovation across algorithms, systems, models, and architectures, SenseTime has achieved dual breakthroughs in cost reduction and efficiency improvement. 

At this month’s SenseTime Product Launch Week, the company unveiled a variety of AI applications and products built on the SenseNova large model. These innovations drive efficiency improvements and intelligent experience upgrades across various scenarios, including video creation, AI-powered office work, e-commerce live streaming, embodied intelligence, and daily life applications, facilitating the transformation of AI from a "productivity tool" to "an integral element of productivity" itself.