Baidu Releases Unlimited OCR, a 3B Model That Keeps the KV Cache Flat for Long-Document Parsing

Baidu has released Unlimited OCR, an open-source 3-billion parameter model designed to process lengthy documents in a single pass. The model utilizes a technique called Reference Sliding Window Attention to keep memory usage and latency stable as document length increases. By maintaining a constant key-value cache size, the architecture improves efficiency for large-scale data extraction tasks compared to standard models that see performance degradation as inputs grow.