According to informed sources, Sun Tianxiang, head of Baidu's Basic Model Unit (BMU), has brought in a researcher known by the alias "Wu Qin" from a Frontier Lab in North America to strengthen the pre-training efforts of the ERNIE large model; meanwhile, Wei Haoran and his team, formerly from DeepSeek, have transferred to BMU, with Wei Haoran assuming the role of lead algorithm developer for ERNIE Multimodal. Wei Haoran was the first author of DeepSeek-OCR and DeepSeek-OCR 2. His departure from DeepSeek had already been noted in the DeepSeek-V4 technical report.

In June, Baidu open-sourced Unlimited OCR, which achieved a record-breaking 93.92% score on the OmniDocBench v1.6 benchmark, setting a new standard for end-to-end OCR performance. The model retains DeepSeek-OCR’s DeepEncoder architecture while re-designing the decoding attention mechanism to enable continuous parsing of dozens of document pages. Sun Tianxiang took over BMU in July and subsequently joined Baidu’s Model Council and Technical Strategy Committee. Within two months, both core tracks—pre-training and multimodal development—began undergoing significant enhancement.