Qianfan OCR
launchMarch 2026
powered byQianfan OCR 4B
goblin vibe check:
handles all your document parsing needs in one go but you'll be dealing with baidu's ecosystem and documentation
document intelligence model from baidu qianfan that handles ocr, layout analysis, table extraction, and document q&a in one pass.
speed
1
page/s
End-to-end OCR plus layout reasoningHandles tables, formulas, charts, and doc QA192-language supportTop OmniDocBench v1.5 score at release
key features
End-to-end OCR plus layout reasoningHandles tables, formulas, charts, and doc QA192-language supportTop OmniDocBench v1.5 score at release
spec & usage
Uses a layout-as-thought step to infer boxes and reading order before generating output
Replaces detect-recognize-assembly pipelines with a single model to cut error propagation
Runs efficiently with vLLM and can process about one page per second on a quantized A100 setup
scope:
datasearchresearchapicloudopen-sourcereal-timefreemultimodal
launchMarch 2026