百度新OCR模型讓50頁文件處理的記憶體成本保持恆定
百度研究人員發布了 unlimited ocr，這是一款能夠在單次推理中處理長達 50 頁文件且記憶體佔用恆定的模型。這一突破消除了通常迫使 ocr 系統逐頁處理長文件的記憶體瓶頸。
baidu introduced unlimited ocr, a solution to the inference memory bottleneck where key-value caches grow linearly with text length. by implementing reference sliding window attention (r-swa), the model attends to all visual tokens but only the last 128 generated tokens, keeping memory use flat. based on deepseek ocr, it uses a 3-billion parameter mixture-of-experts decoder and a frozen deepencoder that compresses 1024x1024 images into 256 tokens. training on 2 million samples showed that this approach achieves 93% on the omnidocbench v1.5 benchmark, proving that constant memory inference does not require sacrificing multi-page recognition quality.