MarkTechPost
6/25/2026

Baidu Releases Unlimited OCR, a 3B Model That Keeps the KV Cache Flat for Long-Document Parsing
Short summary
Baidu open-sourced Unlimited OCR, a 3B-parameter MoE model that processes dozens of document pages in a single pass while keeping KV cache constant via Reference Sliding Window Attention (R-SWA). The architecture maintains flat memory and latency regardless of output length, scoring 93.23 on OmniDocBench v1.5 (6.22 points ahead of DeepSeek). Available under MIT license for immediate integration.
- •3B-parameter MoE model processes multiple document pages in one forward pass
- •Reference Sliding Window Attention keeps KV cache constant, preventing memory/latency bloat
- •Benchmarks at 93.23 on OmniDocBench v1.5, outperforming DeepSeek by 6.22 points; MIT licensed
Generated with AI, which can make mistakes.
Is this a good recommendation for you?



