IT Home On October 29, in response to rumors that Wu Tian, vice president of Baidu Group and deputy director of Baidu Research Institute, had left, Baidu’s public relations manager commented on social media, confirming that Wu Tian had ended her tenure at Baidu. She said that Baidu’s technical work is still progressing.
According to people close to Baidu, there will be new developments regarding the Wenxin model that will be released externally next week. IT Home previously reported that Baidu World 2026 will take place on November 12. Official information indicates that the conference will showcase the technical highlights and application potential of Wenxin Unlimited OCR 2, and will also release the latest upgrades of Baidu’s “Xinyun Model Body” full-stack AI.
Public information shows that Wu Tian worked at Baidu for about 20 years. After graduating with a master's degree from Zhejiang University in 2006, she joined Baidu and started in a technology research position. She was one of the main technical leaders of the early Wenxin Yinian model.
In November 2025, Baidu established a Basic Model R&D department, with Wu Tian serving as the head, reporting directly to CEO Robin Li. Subsequently, the Baidu Basic Model team underwent organizational changes: in May of this year, Wu Tian no longer headed the Basic Model R&D department and was promoted to vice president of Baidu Research Institute; in July, technical expert Sun Tianxiang in the large model field joined Baidu and took over the Basic Model R&D department.
On October 27, Baidu released and made open-sourced Wenxin Unlimited OCR 2, which expands the processing limit of document parsing from dozens of pages to any number of pages. The model achieved combined scores of 84.6 points and 77.1 points in MPDocBench for multi-page documents and MarathonDocBench for hundred-page ultra-long documents, reaching the state-of-the-art end-to-end performance. The technical core of this model is the sparse reference sliding window attention mechanism. In the scenario of hundred-page documents, the single-step decoding speed is 7 times faster than Full Attention and 5.6 times faster than the previous generation solution. In actual testing, this model can parse all 7 works of “Harry Potter” in one go, with a total of over 3600 pages.
Related Reads: