Skip to content
Neural Daily

ModelsNews

Baidu Unlimited OCR 2 parses hundred-page documents

🇭🇺 Olvasd el magyarul →

Baidu released the 3.34B-parameter Unlimited OCR 2 under the MIT license. Its new attention mechanism keeps decoding latency and memory nearly constant, no matter how long the document is.

By: Mikola Máté Patrik · · 2 min read

Unlimited OCR 2 diagram comparing R-SWA visual token overload with SR-SWA human-like selection of visual tokens
Image: Baidu · Press kit · Official image from the Unlimited OCR 2 model card · Source

What was released?

Baidu released Unlimited OCR 2 on October 27 (Hugging Face). The 3.34B-parameter document parsing model is MIT-licensed, and its BF16 weights take about 6.7 GB. The first version came out in June as a follow-up to DeepSeek-OCR.

What is new?

The key is SR-SWA (Sparse Reference Sliding Window Attention). Instead of attending to the full, ever-growing sequence, the model uses a fixed set of about 512 reference tokens: a 128-token local window plus the 16 most relevant 32-token blocks. As a result, per-step latency and memory stay nearly independent of document length. Each page takes 273 image tokens.

How good is it?

On Baidu's own benchmarks, it scores 84.6 on MPDocBench (dozens of pages), ahead of Qwen3.8-Max at 78.3 and Qwen-3.8-Flash-Next at 77.4. On the hundred-page MarathonDocBench it scores 77.1, against 33.7 for Kimi-K3 and 28.6 for Gemini-3.8-Flash. Qwen3.8-Max drops to 22.9 there, and the predecessor scored 0 (MarkTechPost).

Unlimited OCR 2 benchmark chart comparing MarathonDocBench and MPDocBench scores with other models
Baidu · Press kit · Official benchmark chart from the model card · Source

Source: Baidu. MarathonDocBench and MPDocBench results from the model card, self-reported.

Its weakest area is heading hierarchy, at 60.1 and 48.2. Note that all figures are self-reported and MarathonDocBench is not public.

Who should try it?

If you process long contracts, reports or scanned archives, this is one of the most interesting open options right now. It runs on Transformers for quick tests and a custom SGLang build for serving. Benchmark it on your own documents before a larger rollout, since the vendor ran these tests itself.

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles