SenseTime releases “RiRiXin” integrated large model, which is comparable to DeepSeek V3

SenseTime officially launched the "Daily New" integrated large model. This model has achieved significant improvements in multi-modal information processing capabilities and deep reasoning capabilities, and won first place in two authoritative evaluation lists.

According to the "Chinese Large Model Benchmark Evaluation 2024 Annual Report" released by SuperCLUE, an authoritative domestic evaluation agency, SenseTime's "Ririxin" fusion large model tied for the top spot in the country with DeepSeek V3 with a high score of 68.3. In addition, in the multi-modal evaluation of OpenCompass, this model also ranked first, with a score that significantly surpassed GPT-4o.

SenseTime’s “Daily New” large fusion model marks a substantial breakthrough in the field of native fusion modal training. This means that the model can achieve the fusion of text, images, videos and other information, bringing new development directions to the industry. In SuperCLUE's evaluation, the model performed well in both liberal arts and science subjects. Liberal arts ranked first in the world with a score of 81.8, and science subjects also won a gold medal. The computing dimension ranked first in the country with a score of 78.2.

The characteristic of the "RiRiXin" fusion large model is that it can process a variety of complex information at the same time. It can not only identify fuzzy text, but also quickly analyze data charts and provide in-depth reasoning support. In practical application scenarios, this model has shown its unique advantages, especially in fields such as autonomous driving, video interaction, office education, finance, and industrial manufacturing.

During the product development process, SenseTime uses a large amount of interleaved graphics and text data, and uses advanced synthesis technology to establish an interactive bridge between modalities, thereby enhancing the model's ability to understand multi-modal information. In addition, SenseTime has also built a wealth of cross-modal tasks, which provides a solid foundation for the training of "RiRiXin" fusion large models. Through this innovative training method, SenseTime's fusion modal model can effectively respond to user needs in a variety of business scenarios, achieving a virtuous cycle of application implementation and basic model iteration.