Baidu open-sources Unlimited OCR 2, a 3.3B model that parses documents hundreds of pages long

What Baidu released

Baidu released Unlimited OCR 2 on 27 October 2026, with weights on Hugging Face and code on GitHub. The model card lists the MIT licence, and the published weights hold about 3.3 billion parameters. The model reads page images and returns structured text, including tables, formulas, headings, figures and reading order. Baidu frames it as a step from long-horizon to unlimited-horizon document parsing.

It follows the first Unlimited OCR, which Baidu released in June and described as an attempt to push DeepSeek-OCR further. The project's GitHub repository has close to 30,000 stars. The new model is also available on ModelScope and through Baidu Cloud, and a Hugging Face demo is live.

How it handles long documents

The main change is Sparse Reference Sliding Window Attention, or SR-SWA. Baidu says it compresses the interaction between text and visual tokens in very long sequences into a fixed set of working references, for example 512 tokens. That bounds compute, so latency per decoding step and memory stay nearly constant no matter how long the document gets. In practice the model can parse a whole multi-page document in one request and use context across pages, instead of treating each page alone.

For serving, Baidu supports Hugging Face Transformers for quick tests and a customized SGLang build for production. The SGLang setup processes very long documents in page windows, reusing part of the previous window's cache.

Benchmark claims

Baidu also published MarathonDocBench, a benchmark of documents hundreds of pages long, under the Apache 2.0 licence. In Baidu's chart, Unlimited OCR 2 scores 84.6 overall on MPDocBench, which uses documents of dozens of pages, and 77.1 on MarathonDocBench. The best other model shown on MarathonDocBench is Kimi-K3 at 33.7. Qwen3.8-Max scores 78.3 on MPDocBench but 22.9 on MarathonDocBench, and the first Unlimited OCR scores 0.0 there. The gap is smaller on shorter documents. On MPDocBench, Baidu reports 94.8 for plain text and 89.4 for formulas, just ahead of Qwen3.8-Max at 94.3 and 89.2. These are Baidu's own results on a benchmark it created, and independent tests have not yet been published.

What it means and what to watch

Long reports, contracts and filings are where OCR pipelines usually break, because page-by-page parsing loses tables and references that span pages. A small MIT-licensed model that handles them in one pass is easy to self-host. It joins other recent open releases such as Granite Switch 4.1 and LFM3. Watch for independent MarathonDocBench results and support in mainstream vLLM. More open model news is on the Open Source AI hub.

Questions people ask

Is Unlimited OCR 2 free to use commercially?

The model card lists the MIT licence, which allows commercial use. The MarathonDocBench dataset is under Apache 2.0.

How large is Unlimited OCR 2?

About 3.3 billion parameters, according to the safetensors files on Hugging Face.

How long a document can Unlimited OCR 2 parse?

Baidu says SR-SWA keeps memory and latency nearly constant regardless of length, and tests it on documents hundreds of pages long. No hard page limit is stated.

Sources

1. Unlimited OCR 2 model card (Baidu on Hugging Face, 27 Oct 2026)
2. baidu/Unlimited-OCR repository (Baidu on GitHub, 27 Oct 2026)
3. MarathonDocBench dataset (Baidu on Hugging Face, 27 Oct 2026)