Baidu's Foundation Model Unit (BMU) is strengthening its core model team, according to sources familiar with the matter. Sun Tianxiang, who leads the unit, has brought in a researcher codenamed Wuqin from a North American Frontier Lab to enhance ERNIE pre-training. The researcher, who participated in the development of several generations of advanced large models over the past two years, has not been identified by real name or former employer.
Separately, former DeepSeek researcher Wei Haoran and his team have transferred from another part of Baidu to BMU, where Wei serves as head of multimodal algorithms. This move had been hinted at earlier. In June, when Baidu open-sourced Unlimited OCR, one of the three core contributors, a technical director, was credited only as 'YY,' leading the community to speculate it was Wei. Wei is the first author of DeepSeek-OCR and DeepSeek-OCR 2, and the DeepSeek-V4 technical report lists him as having left the company.
Unlimited OCR achieved a score of 93.92% on OmniDocBench v1.6, setting a new record for end-to-end OCR at the time. It uses DeepEncoder from DeepSeek-OCR and redesigned attention mechanisms in the decoder, allowing it to parse dozens of pages of documents in a single pass.
Sun Tianxiang took over BMU in July, simultaneously joining Baidu's Model Committee, and by the end of the month, he had also entered the company's Technical Strategy Committee. Two months later, both the pre-training and multimodal lines of ERNIE are seeing personnel reinforcements.