Baidu's Unlimited-OCR model parses entire pages in one shot, skipping the layout analysis stage

Baidu has released a new vision-language model called Unlimited-OCR that takes a fundamentally different approach to document parsing. Instead of the classic multi-stage pipeline—where you’d run a separate layout analysis model, then crop regions, then OCR each chunk—this 3B-parameter model reads an entire page in a single decoding pass. Headings, paragraphs, tables, everything. One forward pass, one output.

The model is available on Hugging Face under baidu/Unlimited-OCR and weighs in at roughly 6 GB when loaded in bfloat16. The tutorial I walked through sets up the full inference workflow: GPU configuration with automatic dtype selection (bfloat16 if your hardware supports it, float16 if not), dependency installation covering transformers 4.57.1, PyMuPDF, and einops, and then a straightforward model load using trust_remote_code=True. Nothing exotic on the infrastructure side—if you’ve got a Colab GPU, you can run this.

What caught my attention is the two inference modes. The tiled “Gundam” mode handles high-resolution documents by splitting the image into overlapping tiles for fine-grained OCR, while “Base” mode processes the full page at once for faster throughput. The tutorial demonstrates both on synthetic quarterly report pages, then extends the pipeline to multi-page PDFs using PyMuPDF and an infer_multi() function. That’s where things get interesting: the model explicitly preserves cross-page context, so references spanning page breaks don’t get mangled. The generation config also bakes in repetition controls and long-context settings, which tells me the team put real thought into production behavior rather than just benchmark numbers.

I’m not ready to throw out Tesseract and layout parsers just yet—the tutorial uses clean, generated documents with consistent formatting, and I’d want to see this thing handle scanned receipts, rotated text, and the general chaos of real-world PDFs before declaring victory. But the architectural bet here is the right one. If a single model can genuinely absorb the entire visual structure of a page and produce structured text without the fragile hand-offs between detection and recognition stages, that eliminates a huge class of failure modes. The question is whether the 3B parameter budget buys enough robustness for messy inputs.