Baidu's Unlimited-OCR Is a 3B Model That Parses 100-Page PDFs in One Shot: Here's Why It Matters

I saw a YouTube Short the other day that stopped me mid-scroll. China just open-sourced an OCR model that can read an entire 100-page PDF in one shot. And it's only 3 billion parameters. I had to try it.

The model is called Unlimited-OCR, and it's from Baidu. It's a 3B parameter MoE vision-language model built on the DeepSeek-OCR lineage. It's on GitHub with 17K stars and HuggingFace with over 2.2 million downloads already. For context, that's more downloads in a month than most OCR tools get in a year.

What Makes It Different

Every other OCR tool chops your document into individual pages and processes them one at a time. That means you lose cross-page context. Tables that span pages break. Headers and footers in multi-page documents get muddled. It's a mess.

Unlimited-OCR takes the whole thing in one pass. You feed it a PDF, it converts the pages to images internally, and runs them all through a single inference with a 32K token context window. The output is a single coherent document with all the cross-page references intact.

Here's the actual code to run it with transformers:

from transformers import AutoModel, AutoTokenizer

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

# Single image
model.infer(
    tokenizer,
    prompt="<image>document parsing.",
    image_file="page.png",
    output_path="./output",
    base_size=1024, image_size=640, crop_mode=True,
    max_length=32768,
)

# Multi-page PDF
model.infer_multi(
    tokenizer,
    prompt="<image>Multi page parsing.",
    image_files=pdf_to_images("doc.pdf"),
    output_path="./output",
    image_size=1024, max_length=32768,
)

The Numbers Are Impressive

3B parameters is tiny for a vision-language model. For comparison, GPT-4o is estimated at over a trillion parameters. Llama 3.1 is 70B or 405B. This thing runs on a single consumer GPU.

It supports vLLM and SGLang for production deployment. There are already GGUF quantized versions for running on CPU, and AWQ for even faster inference on GPU. The HuggingFace space demo lets you try it right now.

The paper is on arXiv (2606.23050) and the team has been shipping updates fast. Within a week of release, it had support for vLLM, SGLang, ModelScope, and Baidu Cloud.

The Catch

It's not perfect. The 32K context window means you can't do truly massive documents in one shot. A 100-page PDF at 300 DPI generates a lot of image tokens. For anything beyond that, you'd need to chunk it.

The gundam mode (cropped, 640px images) is fast but might miss marginal content. The base mode (1024px, no crop) is slower but more accurate. You need to pick the right config for your use case.

And it's a Baidu model. The documentation is in Chinese-first with English translations that sometimes lag. The community is growing fast though, so that's improving.

What I Think

This is one of those rare releases where the hype is actually underselling it. A 3B model that can read 100 pages in one shot, runs locally, and is open source? That changes what's possible for document processing.

I've been using it to parse scanned contracts and research papers. The cross-page context handling is legitimately better than anything I've used before. It catches table continuations and footnote references that every other tool misses.

If you deal with multi-page documents regularly, go try it. The GitHub repo has everything you need to get started in 10 minutes.

Links

GitHub: baidu/Unlimited-OCR

HuggingFace: baidu/Unlimited-OCR

Paper on arXiv