A lot of models that can transcribe documents have limitations on the number of pages they can do in one pass. Unlimited-OCR is changing that. It is a module that can transcribe dozens pages of documents in a single forward pass. It uses DeepSeek OCR and replace the attention layers in the decoder with Reference Sliding Window Attention. By this approach, it is possible to reduce computation costs while maintaining a KV cache, which is all about speeding up text generation by storing intermediate attention computations rather than recalculating them all the time.
BAIDU JUST DROPPED AN ABSOLUTE GAME-CHANGER FOR DOCUMENT AI
It’s called `Unlimited-OCR`, and it can literally transcribe an entire book in a single pass 🤯
Most vision models read a single page, forget the context, and eventually hit a wall where performance degrades and… pic.twitter.com/KUHrWFHYTW
— Charly Wargnier (@DataChaz) June 24, 2026
[HT]
