2025 DeepSeek-OCR技术报告： 视觉压缩长文本的探索性研究.pdfVIP

DeepSeek-OCR:ContextsOpticalCompression

HaoranWei,YaofengSun,YukunLi

DeepSeek-AI

Abstract

WepresentDeepSeek-OCRasaninitialinvestigationintothefeasibilityofcompressinglong

contextsviaoptical2Dmapping.DeepSeek-OCRconsistsoftwocomponents:DeepEncoder

andDeepSeek3B-MoE-A570Masthedecoder.Specifically,DeepEncoderservesasthecore

engine,designedtomaintainlowactivationsunderhigh-resolutioninputwhileachievinghigh

compressionratiostoensureanoptimalandmanageablenumberofvisiontokens.Experiments

showthatwhenthenumberoftexttokensiswithin10timesthatofvisiontokens(i.e.,a

compressionratio10×),themodelcanachievedecoding(OCR)precisionof97%.Evenata

compressionratioof20×,theOCRaccuracystillremainsatabout60%.Thisshowsconsiderable

promiseforresearchareassuchashistoricallong-contextcompressionandmemoryforgetting

mechanismsinLLMs.Beyondthis,DeepSeek-OCRalsodemonstrateshighpracticalvalue.

OnOmniDocBench,itsurpassesGOT-OCR2.0(256tokens/page)usingonly100visiontokens,

andoutperformsMinerU2.0(6000+tokensperpageonaverage)whileutilizingfewerthan

800visiontokens.Inproduction,DeepSeek-OCRcangeneratetrainingdataforLLMs/VLMs

atascaleof200k+pagesperday(asingleA100-40G).Codesandmodelweightsarepublicly

accessibleat/deepseek-ai/DeepSeek-OCR.

0.1DeepSeek-OCR

64

vis

toks(left)100

vis

toks(left)64

vis

toks(right)100

vis

toks(right)(Gundam-M200dpi)