Model WarsJune 25, 2026via MarkTechPost

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

Why it matters

Baidu's Unlimited OCR demonstrates a practical breakthrough in long-context efficiency through novel attention mechanisms (R-SWA), directly challenging DeepSeek's OCR capabilities and showing how MoE + architectural innovation can solve a real enterprise problem—document processing at scale—without the compute tax.

Key signals

  • 3B-parameter MoE model
  • Processes dozens of document pages in single forward pass
  • Reference Sliding Window Attention (R-SWA) maintains flat KV cache
  • Memory and latency scale constant with output growth
  • Scores 93.23 on OmniDocBench v1.5
  • Beats DeepSeek OCR baseline by 6.22 points
  • MIT license (open-source)
  • Published by Baidu

The hook

93.23 on OmniDocBench. Baidu just open-sourced a 3B model that processes 50+ document pages without memory overhead.

Baidu open-sourced Unlimited OCR, a 3B-parameter MoE model that parses dozens of document pages in a single forward pass. Its Reference Sliding Window Attention (R-SWA) holds the KV cache constant, so memory and latency stay flat as output grows. It scores 93.23 on OmniDocBench v1.5, beating the DeepSeek OCR baseline by 6.22 points, under an MIT license.

Get stories like this every Friday.

The 5 AI stories that matter — free, in your inbox.

Free forever. No spam.

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