88Z AI, 23 June — Baidu has released the open-source OCR model Unlimited OCR, specializing in long document parsing. It achieved a new state-of-the-art comprehensive score of 93.23% on the OmniDocBench benchmark, surpassing DeepSeek OCR. The core innovation is the Reference Sliding Window Attention mechanism, which uses a soft forgetting strategy to maintain constant KV Cache, ensuring inference speed does not drop as document length increases, with TPS improving by approximately 35% at 6000 tokens.
At a glance
- Unlimited OCR sets a new SOTA on OmniDocBench with a 93.23% score, surpassing DeepSeek OCR.
- Uses Reference Sliding Window Attention mechanism, improving TPS by approximately 35% at 6000 tokens.
The update
Baidu has officially open-sourced Unlimited OCR, a new end-to-end long document OCR model. Designed to parse dozens of pages in a single pass, the model achieved a breakthrough on the OmniDocBench benchmark with a comprehensive score of 93.23%, setting a new state-of-the-art and outperforming DeepSeek OCR. Technologically, the core innovation of Unlimited OCR lies in its Reference Sliding Window Attention mechanism. By implementing a soft forgetting strategy, the model maintains a constant KV Cache, ensuring that inference speed does not degrade as document length increases. Test data shows that when processing 6000 tokens, its tokens per second improved by approximately 35%, significantly optimizing efficiency for long text processing.
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.