Nate@nate_512

ติดตาม

Older models choke on long documents, losing context page by page while inference speeds plummet as the text grows. Baidu Inc. just released Unlimited-OCR to solve this mess, building a new document AI model on top of DeepSeek OCR that actually fixes the memory bloat issue. The trick involves modifying the attention mechanism to use Reference Sliding Window Attention instead of trying to hold everything in active RAM. Each token only accesses the current page and the previous 128 words, creating a design where the memory footprint stays fixed regardless of how massive the output gets. This architecture ensures generation speed remains steady even when processing dozens of pages in a single pass without the usual slowdowns. Performance numbers back up the hype: Unlimited-OCR hits a 93% score on standard parsing benchmarks, beating the older baseline by six points. Error rates stay below 0.11 even when pushing past 40 pages, avoiding the 35% slowdown that used to cripple older models at scale. The whole thing is free and open-source with the repository, weights, and paper available via a linked thread. It frames the industry shift toward one-shot long-horizon parsing where you can transcribe an entire book without hitting a wall.

5d

Nate@nate_512

ติดตาม

Older models choke on long documents, losing context page by page while inference speeds plummet as the text grows. Baidu Inc. just released Unlimited-OCR to solve this mess, building a new document AI model on top of DeepSeek OCR that actually fixes the memory bloat issue. The trick involves modifying the attention mechanism to use Reference Sliding Window Attention instead of trying to hold everything in active RAM. Each token only accesses the current page and the previous 128 words, creating a design where the memory footprint stays fixed regardless of how massive the output gets. This architecture ensures generation speed remains steady even when processing dozens of pages in a single pass without the usual slowdowns. Performance numbers back up the hype: Unlimited-OCR hits a 93% score on standard parsing benchmarks, beating the older baseline by six points. Error rates stay below 0.11 even when pushing past 40 pages, avoiding the 35% slowdown that used to cripple older models at scale. The whole thing is free and open-source with the repository, weights, and paper available via a linked thread. It frames the industry shift toward one-shot long-horizon parsing where you can transcribe an entire book without hitting a wall.

5d