Research Vault

Baidu Unlimited-OCR

Properties1
tagshuggingface, image-text-to-text, baidu, vision-language, ocr, document-parsing, pdf-parsing, multilingual, deepseek-v2, custom-code

2 min read

Baidu Unlimited-OCR

One-shot long-horizon document parsing — vision-language OCR model from Baidu (PaddlePaddle team). Aimed at pushing DeepSeek-OCR further by extending “one-shot” parsing to longer documents and multi-page inputs.

What it does

  • Single-image document parsing (two configs: gundam and base)
  • Multi-page document parsing across multiple images in a single forward pass
  • PDF parsing via PDF→image conversion + multi-page mode
  • Schema-constrained decoding (via the DeepSeek-OCR no-repeat-n-gram logit processor) to prevent output loops on long sequences
  • Max sequence length: 32,768 tokens

Architecture

Custom transformers model based on the DeepSeek V2 backbone (configuration_deepseek_v2.py in the repo). The repo ships:

  • Vision encoder (deepencoder.py)
  • Multi-modal model (modeling_unlimitedocr.py, modeling_deepseekv2.py)
  • Conversation template (conversation.py)

The model is implemented with trust_remote_code — you must opt in to load it. Tested on python 3.12.3 + CUDA 12.9 with torch==2.10.0, transformers==4.57.1.

Release timeline

DateEvent
2026-06-22Initial release on Hugging Face and GitHub
2026-06-23Paper posted to arXiv (2606.23050); mirrored on ModelScope as PaddlePaddle/Unlimited-OCR
2026-06-24HF Spaces demo live (courtesy of AK)

Notable tags / metadata

  • Pipeline: image-text-to-text
  • License: MIT (open-weight, commercial-friendly)
  • Languages: multilingual
  • Architecture: DeepSeek V2 backbone with custom vision encoder and OCR-specific decoding
  • Use case: long-horizon OCR / document parsing — addresses the “one-shot” limit of prior OCR models

How to use

from transformers import AutoModel, AutoTokenizer
import torch
 
model_name = 'baidu/Unlimited-OCR'
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_name,
    trust_remote_code=True,
    use_safetensors=True,
    torch_dtype=torch.bfloat16,
).eval().cuda()
 
# Single image, gundam config
model.infer(tokenizer, prompt='<image>document parsing.',
            image_file='your_image.jpg', base_size=1024,
            image_size=640, crop_mode=True, max_length=32768,
            no_repeat_ngram_size=35, ngram_window=128, save_results=True)

The model also ships an SGLang server recipe for production serving.

Benchmarks

No quantitative benchmark tables are published in the model card. The repo points to the arXiv paper (“Unlimited OCR Works”) for evaluation details. The model’s value proposition — long-horizon one-shot parsing — is showcased in the assets/long-horizon-ocr.gif demo in the repo.

See also

Graph View