Unlimited-OCR
Baidu's 3.3B MIT-licensed OCR model parses documents of dozens to 100+ pages in one pass, with bounded attention memory.
In short. Unlimited-OCR turns scanned pages and PDFs into text. Most OCR models that use a language model slow down and use more memory as their output grows; this one is built to keep both steady, so a long document can be read in one go. It is small, at 3.3 billion parameters, and MIT licensed.
Summary
Baidu released Unlimited-OCR on 22 June 2026 as a follow-up to DeepSeek-OCR, and Unlimited-OCR 2 on 27 October 2026. Both have about 3.3 billion parameters, ship in BF16 at about 6.7 GB, and use the same kind of mixture-of-experts decoder as DeepSeek-OCR, with 64 routed experts, 6 chosen per token and 2 shared. Both are released under the MIT license. Baidu's reasoning is that a language-model decoder gets slower and heavier as its output grows, because its KV cache keeps accumulating, while a person copying a long text does not slow down.
The first model replaces every attention layer in the decoder with Reference Sliding Window Attention, which keeps the KV cache constant while decoding. Baidu says it can transcribe dozens of pages in a single forward pass within a 32K length limit. Unlimited-OCR 2 goes further with Sparse Reference Sliding Window Attention, which limits attention to a fixed set of working references, for example 512 tokens, so each decoding step costs nearly the same however long the document gets. Baidu reports 7.0 and 5.6 times faster decoding steps than full attention and the first design at hundred-page scale, and top results on MPDocBench and on MarathonDocBench, a hundred-page benchmark that Baidu created for the purpose. All of these figures come from Baidu's own reports.
The model reads page images, so a PDF is first converted to images; the examples render pages at 300 dpi and send them together so the model can use context across pages. Prompts such as "document parsing." for a single page and "Multi page parsing." for several pages are documented. It runs with Transformers using custom code, with SGLang (recommended for Unlimited-OCR 2) and, for the first release, with vLLM. Baidu tested Unlimited-OCR 2 on an 80 GB A800 and does not state a minimum memory. The first repository stores the weights twice, so the real download is about 6.7 GB.
Good for
- Converting long PDFs and multi-page scans to Markdown in one request, with context across pages
- High-volume document digitization on a single GPU, since the model is small
- Researchers studying decoders whose memory use stays constant as the output grows
- Commercial use, since the weights are MIT licensed
Watch out for
- The two releases are different designs, and the recommended server differs, so test the one you plan to deploy
- It needs trust_remote_code and specific library versions, so pin your environment
- The MarathonDocBench results come from a benchmark that Baidu created and reported itself
- Raw output includes layout markers that you must strip for some benchmarks and pipelines
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Unlimited-OCR-2 variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Unlimited-OCR-2 BF16 | 3.34B | 7.68 GB | 3.84 GB | 2.3 GB | Laptop or phone |
| Unlimited-OCR BF16 | 3.34B | 7.68 GB | 3.84 GB | 2.3 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
- Best open-source OCR and document models in 2026
- Eleven open OCR models in a year: small and hard to compare
Frequently asked questions
- What is Unlimited-OCR?
- Unlimited-OCR turns scanned pages and PDFs into text. Most OCR models that use a language model slow down and use more memory as their output grows; this one is built to keep both steady, so a long document can be read in one go. It is small, at 3.3 billion parameters, and MIT licensed.
- Can I use Unlimited-OCR commercially?
- Yes. Unlimited-OCR is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Unlimited-OCR open source?
- Unlimited-OCR is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Unlimited-OCR on my own computer?
- Yes. The smallest variant, Unlimited-OCR-2, needs about 2.3 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Unlimited-OCR released?
- Unlimited-OCR was released by Baidu on June 22, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Unlimited-OCR-2 model card (Hugging Face) huggingface.co
- Unlimited OCR Works 2 (technical report) arxiv.org
- Unlimited-OCR model card (Hugging Face) huggingface.co
- Unlimited OCR Works (technical report) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| DeepSeek-OCR 2 DeepSeek | 3.39B | – | Apache 2.0 (Permissive) | 2.34 GB Laptop or phone | |
| GLM-OCR Z.ai (Zhipu AI) | 900M | – | MIT (Permissive) | 0.621 GB Laptop or phone | |
| PaddleOCR-VL Baidu | 960M | – | Apache 2.0 (Permissive) | 0.662 GB Laptop or phone | |
| ERNIE 5.5 Flash Base Baidu | 280B 14B active | 1M | ERNIE 5.5 Flash Base (Conditional) | 193 GB Big-memory desktop |