Today for AI
HOT RADAR
llmHEAT 0.0°

AI CURATED HOT STORY · 10/27/2026

Baidu Open-Sources Unlimited OCR 2, Cutting Long-Doc Memory by 96%

4 reports archived0 independent sourcesupdated 10/27/2026, 16:35:33
Synthesis & Latest Updates

Baidu has officially open-sourced Wenxin Unlimited OCR 2, a 3.3-billion-parameter lightweight model that achieves a breakthrough in long-document parsing via the Sparse Reference Sliding Window Attention (SR-SWA) mechanism. This innovation enables the continuous processing of hundreds or even thousands of pages in a single request while maintaining nearly constant decoding latency and memory usage.

Cross-verified reports highlight the model's extreme resource efficiency: peak KV cache usage drops by 96% from 47GB to 1.77GB, with only 500M parameters active during inference. On benchmarks like MarathonDocBench, it scores 77.1, significantly outperforming competitors such as Kimi-K3 and Qwen3.8-Max. Although it shows weaknesses in heading hierarchy recognition and relies on self-reported data, its ability to maintain context coherence over ultra-long documents has been confirmed by community tests.

This development drastically lowers the technical barriers and hardware costs for handling massive document volumes. With full availability on GitHub and Hugging Face, the industry can rapidly deploy efficient, low-cost digitization solutions, accelerating AI adoption in knowledge management, legal compliance, and academic publishing.

LATEST/Wenxin Unlimited OCR 2 is now fully open-source on GitHub and Hugging Face. Community tests confirm it can process thousands of pages, such as the entire Harry Potter series, without interruption, marking a new era of low-cost, high-concurrency long-document AI.

Hourly heat curve

55 fully observed hours
Now
0.0
Peak
48.619:00
24h change
-100%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 4 reports · Latest first
  3. Marktechpost (@Marktechpost)T2·85 pts

    Baidu Open-Sources Unlimited OCR 2 for Long Document Parsing

    • Utilizes SR-SWA to attend to a fixed set of working references (e.g., 512 tokens), keeping per-step decoding latency and memory nearly constant regardless of document length.
    • Scores 77.1 on MarathonDocBench (hundreds of pages), significantly surpassing Kimi-K3 (33.7) and Qwen3.8-Max (22.9); also leads Qwen3.8-Max on MPDocBench.
    • Trade-offs exist: heading hierarchy recognition is its weakest metric (48.2) on long documents, the reference pool is lossy, and all scores are self-reported.
  4. 新智元 (微信公众号)T2·78 pts

    Baidu Open-Sources ERNIE Unlimited OCR 2, Tops Global Long-Document Benchmarks

    • ERNIE Unlimited OCR 2 has 3B total parameters but activates only 500M during inference, balancing performance and efficiency.
    • It achieved top rankings across all metrics in both MarathonDocBench (long-document) and MPDocBench (multi-page/cross-page).
    • Breaks traditional pagination limits by supporting the continuous parsing of ultra-long documents (e.g., entire Harry Potter series) without manual splitting.