What licence is Unlimited OCR 2 under?

MIT, according to the Hugging Face model card. The MarathonDocBench benchmark released with it is under Apache 2.0.

What is new in Unlimited OCR 2?

Sparse Reference Sliding Window Attention, which compresses attention to a fixed set of working references so the model can parse very long documents with nearly constant memory and latency.

Unlimited OCR 2 news

Document parsing vision-language model (open weights)1 story

Unlimited OCR 2 is an open-weight document parsing model from Baidu, released on 27 October 2026 under the MIT licence.

Baidu's long, wave-shaped glass headquarters building in Beijing behind a lawn and trees. — Baidu open-sources Unlimited OCR 2, a 3.3B model that parses documents hundreds of pages long

Open Source AI27 Oct 2026, 09:00 UTC

Baidu released Unlimited OCR 2 under MIT on 27 October 2026. It scores 77.1 on Baidu's new hundred-page MarathonDocBench, where rivals score under 34.

Summary

Baidu released Unlimited OCR 2 on 27 October 2026, an open-weight document parsing model of about 3.3 billion parameters under the MIT licence. A new attention method keeps memory and decoding speed nearly flat as documents grow, so the model can parse documents hundreds of pages long in one pass. On Baidu's new MarathonDocBench it scores 77.1, against 33.7 for the best model Baidu compared it with.

About Unlimited OCR 2

Unlimited OCR 2 is an open-weight document parsing model from Baidu, released on 27 October 2026 under the MIT licence. It converts page images of long documents into structured text and uses Sparse Reference Sliding Window Attention to keep memory and per-step latency nearly constant, so it can parse documents of hundreds of pages in one pass.

Reviewed 28 Oct 2026 · by Baidu

Facts at a glance

Developer: Baidu
Released: 27 October 2026 source
Licence: MIT source
Parameters: About 3.3 billion (BF16 safetensors) source
MPDocBench overall: 84.6, per Baidu's chart source
MarathonDocBench overall: 77.1, per Baidu's chart source
Serving: Transformers and a customized SGLang build source

Questions people ask

What licence is Unlimited OCR 2 under?: MIT, according to the Hugging Face model card. The MarathonDocBench benchmark released with it is under Apache 2.0.
What is new in Unlimited OCR 2?: Sparse Reference Sliding Window Attention, which compresses attention to a fixed set of working references so the model can parse very long documents with nearly constant memory and latency.

Unlimited OCR 2 news

Every story about Unlimited OCR 2

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