Baidu Unlimited-OCR

• URL: https://github.com/baidu/Unlimited-OCR
• Date Saved: 2026-06-27
• Source: GitHub
• Tags: ai-tools, ai-engineering

Summary

百度开源的文档OCR模型，主打「一次性长文档解析」— 单次推理即可解析多页文档。明确表示是在 DeepSeek-OCR 基础上更进一步。

Key Features:

• Single-image OCR: two modes — “gundam” (crop-based, 640px) and “base” (full image, 1024px)
• Multi-page document parsing (multiple images in one inference call)
• Full PDF support (converts pages → images → multi-page parsing)
• Streaming output via OpenAI-compatible API
• Custom n-gram repetition prevention (no_repeat_ngram_size=35)
• Available on HuggingFace + ModelScope

Tech Stack:

• Python / HuggingFace Transformers (trust_remote_code)
• SGLang for high-performance serving
• bfloat16, CUDA 12.9, context length 32,768 tokens
• PyTorch 2.10, transformers 4.57.1
• License: MIT

Notable:

• Created June 18, 2026 — already 10,338 stars, 789 forks (viral)
• arXiv paper: 2606.23050 (17 authors from Baidu)
• Acknowledges DeepSeek-OCR, DeepSeek-OCR-2, PaddleOCR
• HuggingFace Spaces demo available
• PaddleOCR lineage = excellent CJK character support
• Includes custom SGLang wheel in repo for optimized serving