Unlimited OCR
Turn images and PDFs into text for free, with no signup and no page limit. Everything runs in your browser, so your files are never uploaded.
Engine: Tesseract 5 via tesseract.js, running on your device. This is not Baidu's Unlimited-OCR model, which needs a GPU (see below).
What is Unlimited-OCR?
Unlimited-OCR is Baidu's open-source document OCR model, released on 22 June 2026 on GitHub and Hugging Face. It builds on DeepSeek-OCR and reads long, multi-page documents into structured text in a single pass.
| Released | 22 June 2026 by Baidu |
|---|---|
| Licence | MIT (code and weights) |
| Size | 3.34B parameters, BF16; mixture-of-experts decoder with about 500M active parameters |
| Key idea | Reference Sliding Window Attention (R-SWA): the KV cache stays constant, so GPU memory does not grow with document length |
| Long documents | “Dozens of pages” in one forward pass within a 32K output length; evaluated on 2 to 40+ page documents |
| Accuracy | OmniDocBench overall 93.23 (v1.5) and 93.92 (v1.6), reported by the authors |
| Runs on | NVIDIA GPU via Transformers, vLLM or SGLang; community GGUF builds exist |
| Languages | Listed only as “multilingual”; no official language list |
Sources: official README, paper (arXiv 2606.23050), Hugging Face model card. Checked 16 October 2026.
Can I try Unlimited-OCR online?
Yes: Baidu's README links a free demo on Hugging Face Spaces that runs the real model. It uses shared ZeroGPU hardware, so expect a queue and a per-user GPU time quota; sign in to Hugging Face for a larger quota.
Use the demo when you want Unlimited-OCR's layout-aware output on a few pages. Use the tool above when the document is private, when you need many files, or when plain text is enough.
Is Unlimited-OCR free?
Yes, the model is free: Baidu publishes it under the standard MIT licence, which allows commercial use if you keep the copyright notice. The cost is hardware, since it needs an NVIDIA GPU; Baidu also offers it as a hosted API on Baidu Cloud.
How do I run Unlimited-OCR locally?
Install the pinned PyTorch and Transformers versions, load baidu/Unlimited-OCR with trust_remote_code=True, and call model.infer() on an image or model.infer_multi() on PDF pages. For throughput, serve it with vLLM or SGLang.
Step-by-step guide with the official commands and hardware notes →
Unlimited-OCR or in-browser OCR: which should I use?
| Unlimited-OCR (Baidu) | This tool (Tesseract) | |
|---|---|---|
| Best for | Long, complex documents: tables, formulas, multi-column layouts | Clean scans, screenshots, receipts, quick text |
| Where it runs | Your GPU, a cloud GPU, or the HF demo | Your browser, any device |
| Privacy | Private if self-hosted; the demo uploads to Hugging Face | Nothing uploaded |
| Output | Structured text with layout tags | Plain text |
| Cost | Free model, paid hardware | Free |
FAQ
Is this site made by Baidu?
No. Maestro OCR is independent and not affiliated with or endorsed by Baidu. “Unlimited-OCR” is used here only to describe Baidu's open-source model.
Does this OCR tool upload my files?
No. The engine, its WebAssembly core and the language data are served from this site and run in your browser, so files and text stay on your device. You can check this in your browser's network panel.
What does “unlimited” mean here?
For Baidu's model, it means long-horizon parsing with flat memory, so many pages fit in one pass. For this tool, it means no signup, no page cap and no daily limit; only your device's speed limits it.
Why is the first run slower?
The first run downloads the OCR engine (about 4 MB) and the language file (1 to 3 MB). Your browser caches them, so later runs start immediately.