Baidu releases Unlimited OCR: 3B-parameter MoE model with constant KV cache for efficient long-document parsing
Original title:Baidu Releases Unlimited OCR, a 3B Model That Keeps the KV Cache Flat for Long-Document Parsing
AI summary
Baidu launched Unlimited OCR, a 3B-parameter MoE model that activates only 500M parameters during inference. Its core innovation, Reference Sliding Window Attention (R-SWA), fixes the KV cache size at Lm + n (n defaults to 128), so memory and latency do not grow with output length. The model continues training from DeepSeek OCR for 4,000 steps, supports a maximum length of 32K, and achieves 16x token compression through DeepEncoder. It scores 93.23 overall on OmniDocBench v1.5, 6.22 points above the DeepSeek OCR baseline; v1.6 scores 93.92, the highest. In Base mode, throughput reaches 5,580 TPS, 12.7% higher than DeepSeek OCR, and latency is 35% lower at 6,000-token output. Suitable for scenarios such as whole-book transcription, with code and weights open-sourced on HuggingFace.
Key Takeaways
- Baidu launched Unlimited OCR,a 3B-parameter MoE model that activates only 500M parameters during…
- Its core innovation,Reference Sliding Window Attention (R-SWA),fixes the KV cache size at Lm + n…
- The model continues training from DeepSeek OCR for 4,000 steps,supports a maximum length of…
- It scores 93.
- 23 overall on OmniDocBench v1.
Quick Facts
- Published
- Updated
- Source
- MarkTechPost(RSS)
- Category
- Model Releases
- AI score
- 73
Baidu这个OCR模型用R-SWA把KV缓存压成常量,长文档解析终于不用越跑越慢了。MIT开源,3B总参但推理只消500M,做文档管线的可以直接接。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: Baidu releases Unlimited OCR: 3B-parameter MoE model with constant KV cache for efficient long-document parsing. June 25, 2026. https://new.nat6.net/en/news/cmqt3kfx30024sl0ev7aty5q7/