Baidu releases two major models, benchmarking DeepSeek and focusing on multimodal approaches

According to China Business News, on March 16, two years after the official release of ERNIE Bot, Baidu released the multimodal big model Wenxin 4.5 and the benchmark DeepSeek Wenxin X1. Wenxin big model 4.5 is launched on the Baidu AI Cloud Qianfan big model platform, and the input price is 0.004 yuan/thousand tokens; The input price of Wenxin Big Model X1 is 0.002 yuan/thousand tokens, which is half of DeepSeek R1. In addition, Robin Lee disclosed in his internal speech this year that Baidu will release the 5.0 version of Wenxin model in the second half of the year and increase the commercialization of AI applications. Next, Baidu will officially open source the Wenxin big model on June 30th. Through Baidu's native multimodal big model Wenxin 4.5, users can upload files including documents, images, audio, and videos for AI to interpret. The Wenxin Big Model X1 is a deep thinking model that utilizes key technologies such as progressive reinforcement learning, end-to-end training based on thought and action chains, and a diverse and unified reward system. Baidu claims that it benchmarks DeepSeeker R1 in performance and has a "long thought chain". The reporter's experience shows that in addition to online search, it has added multimodal capabilities and multi tool calling, which can understand and generate images, call tools to generate code, charts, etc. Currently, DeepSeeker R1 only recognizes text for uploaded attachments.