88Z AI, 28 October — Baidu Wenxin has released and open-sourced Unlimited OCR 2, an end-to-end document parsing model capable of continuously recognizing documents exceeding 3,600 pages. Its core SR-SWA attention mechanism, combined with chunked decoding, reduces peak KV cache for hundred-page tests by approximately 96% and increases 16-way concurrent throughput by 3.65 times.
At a glance
- Unlimited OCR 2 supports continuous recognition of 3,600+ page documents and is open-source
- SR-SWA mechanism reduces KV cache by ~96% and boosts concurrent throughput by 3.65x
The update
Baidu Wenxin has officially launched and open-sourced Unlimited OCR 2, an end-to-end document parsing model. This model is capable of handling ultra-long documents, continuously recognizing the entire Harry Potter series which contains over 3,600 pages. Technically, Unlimited OCR 2 employs a core SR-SWA attention mechanism that simulates human transcription behavior by filtering relevant pages and information blocks on demand. Combined with a chunked decoding mechanism, the model significantly optimizes resource efficiency. In hundred-page tests, its peak KV cache dropped substantially from 47GB to 1.77GB, a reduction of approximately 96%. Additionally, the system's 16-way concurrent throughput increased to 3.65 times its previous level.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 量子位Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.