Overview

Unlimited OCR is baidu’s open-weight end-to-end document parsing model using reference-sliding-window-attention for constant-memory multi-page OCR. Released June 2026 with weights on Hugging Face.

Specifications

  • Parameters: 3B MoE, 500M active
  • Benchmarks: 93.23% OmniDocBench v1.5 (+6.22 vs DeepSeek OCR baseline); 93.92% v1.6
  • Throughput: 5,580 tokens/second (Base mode)
  • Context: Dozens of pages in single 32K pass
  • License: MIT; GitHub: baidu/Unlimited-OCR; vLLM/SGLang support

Sources