Baidu releases two large models, benchmarking DeepSeek and focusing on multimodality

PANews reported on March 17, according to the First Financial Report, that on March 16, two weeks after the official release of Wenxin Yanyan, Baidu released multi-modal large models Wenxin 4.5 and Wenxin X1 benchmarked against DeepSeek. The large model Wenxin 4.5 was launched on Baidu's Intelligent Cloud Qianfan Large Model Platform with an input price of RMB 0.004 per thousand tokens; the input price for the large model Wenxin X1 is RMB 0.002 per thousand tokens, half that of DeepSeek R1. In addition, Li Yanhong revealed in his internal speech this year that Baidu will release version 5.0 of its large model in the second half of this year and increase commercialization practices for AI applications. Next up, Baidu will officially open source its big model on June 30.

Through Baidu's native multi-modal big model Wenxin 4.5, users can upload files including documents, images, audio and video for AI interpretation.The big model X1 is a deep thinking model using key technologies such as progressive reinforcement learning,end-to-end training based on thought chains and action chains,a unified reward system etc.,Baidu claims it benchmarks against DeepSeek-R1 in performance with "long thought chain". As observed by reporters,in addition to internet search,it has added multimodal capabilities and multiple tool calls,capable of understanding and generating pictures,and calling tools to generate code,charts etc.Currently,the attachments supported by DeepSeek-R1 only recognize text.