Two Key Generals from Baidu MindSpore Join Forces: North America's Frontier Lab + DeepSeek OCR Core Author

Perceiving AI News Flash, Baidu's Basic Model Development Department (BMU) is reinforcing the Core Model Team. Sources said that the head, Sun Tianxiang, has brought in a researcher nicknamed "Wu Qin" from a Frontier Lab in North America to strengthen the Wenxin pre-training with the team. The individual has been involved in the development of several cutting-edge large models over the past two years, but their real name and original unit have not been disclosed. On the other hand, former DeepSeek researcher Wei Haoran and his team have also transferred from within Baidu to BMU, with him serving as the head of the Wenxin multimodal algorithm.

Wei Haoran's involvement actually had foreshadowing. When Baidu open-sourced Unlimited OCR in June, the technical director among the three core contributors only signed as "YY," and the community at the time speculated that it was Wei Haoran. Wei Haoran is the first author of DeepSeek-OCR and DeepSeek-OCR 2; the DeepSeek-V4 technical report also indicated that he had already left the company.

When Unlimited OCR was released, it achieved 93.92% on OmniDocBench v1.6, briefly setting a new end-to-end OCR record. It inherited the DeepEncoder from DeepSeek-OCR and redesigned the decoding end's attention mechanism, enabling continuous parsing of dozens of pages of documents at once.

In July, Sun Tianxiang took over BMU, joined Baidu's Model Committee, and by the end of the month, entered Baidu's Technical Strategy Committee. Two months later, both core lines of Wenxin's pre-training and multimodal capabilities started to onboard new team members.