Baidu releases Qianfan-OCR: a single 4B-parameter model that replaces document processing pipelines

Baidu released Qianfan-OCR, a 4-billion-parameter vision-language model for document intelligence. The model handles document parsing, table extraction, formula recognition, chart understanding, layout analysis, and key information extraction in a single inference pass. Models are available now on HuggingFace (baidu/qianfan-vl). A research paper is on arxiv (2603.13398).

Document intelligence workflows typically chain multiple specialized models: one for layout, one for tables, one for formulas. Qianfan-OCR collapses this into a single 4B-parameter model, reducing latency, integration complexity, and cost. At 4B parameters it is deployable on local hardware rather than requiring cloud inference. For builders working with structured documents — financial reports, technical papers, legal filings, scientific literature — this is a direct alternative to multi-step pipelines using larger models, available today via HuggingFace.