Unlimited-OCR: Baidu’s 3B-Parameter Open-Source OCR That Handles 100-Page PDFs in One Pass

Background

Unlimited-OCR, released by Baidu’s research team in June 2026, targets “one‑shot long‑horizon parsing” so that a model can ingest an entire document (e.g., a 100‑page PDF) in a single forward pass.

Compared with cloud OCR services (AWS Textract, Google Vision, Azure Document Intelligence) that charge $1.5–15 per 1 000 pages, Unlimited-OCR runs locally on a GPU, eliminating per‑page fees and keeping data offline.

Core capabilities

One‑shot long‑text parsing : 32 K token context window enables processing of a 100‑page PDF without page‑wise fragmentation.

Accuracy : Standard benchmark score 93 % (6 % above baseline) and error rate < 0.11 % after 40 pages.

Native multilingual support : Chinese, English, Japanese, Korean and other major languages work out‑of‑the‑box.

Fully offline operation : GPU‑only inference, no network traffic.

Compact model size : 3 B parameters; INT4‑quantized version runs on consumer‑grade GPUs such as RTX 3090.

Framework coverage : Compatible with Transformers, vLLM, SGLang, Ollama, llama.cpp and Docker.

Direct PDF handling : Built‑in PyMuPDF pipeline converts PDF pages to images and runs OCR in a single flow.

OpenAI‑compatible API : Deploy via vLLM or SGLang and call the model through the standard ChatGPT endpoint.

System requirements

GPU memory: ≥ 16 GB for bf16 model; ≥ 8 GB for INT4‑quantized model.

GPU: NVIDIA with CUDA 12.x.

Python: 3.12 or newer.

Dependencies: torch 2.10, transformers 4.57, PyMuPDF, Pillow, einops, etc.

Quick installation (Transformers)

pip install torch torchvision transformers Pillow einops pymupdf

python -c "
import os, torch
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained(
    'baidu/Unlimited-OCR',
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)
model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('baidu/Unlimited-OCR', trust_remote_code=True)

# Single‑image OCR
model.infer(tokenizer,
    prompt='<image>document parsing.',
    image_file='your_image.jpg',
    output_path='./outputs',
    base_size=1024, image_size=640, crop_mode=True,
    max_length=32768,
    no_repeat_ngram_size=35, ngram_window=128,
    save_results=True)

# Multi‑page PDF OCR (images pre‑converted with PyMuPDF)
model.infer_multi(tokenizer,
    prompt='<image>Multi page parsing.',
    image_files=['page1.png', 'page2.png'],
    output_path='./outputs',
    image_size=1024, max_length=32768,
    no_repeat_ngram_size=35, ngram_window=1024,
    save_results=True)
"

Docker one‑click deployment (vLLM)

docker run --rm --gpus all --network host --ipc host \
  vllm/vllm-openai:unlimited-ocr \
  baidu/Unlimited-OCR \
  --trust-remote-code \
  --logits_processors vllm.model_executor.models.unlimited_ocr:NGramPerReqLogitProcessor \
  --no-enable-prefix-caching \
  --mm-processor-cache-gb 0

After the container starts, the model is reachable via the OpenAI‑compatible endpoint:

POST http://127.0.0.1:8000/v1/chat/completions

Related resources

GitHub: https://github.com/baidu/Unlimited-OCR

Hugging Face: https://huggingface.co/baidu/Unlimited-OCR

arXiv paper: https://arxiv.org/abs/2606.23050

vLLM recipe: https://recipes.vllm.ai/baidu/Unlimited-OCR

Online demo: https://huggingface.co/spaces/baidu/Unlimited-OCR