Originally published on rohitraj.tech
Baidu open-sourced Unlimited-OCR on June 22, 2026 (MIT) — a 3B mixture-of-experts model with 500M active params that parses 40+ page documents in a single forward pass. Its new Reference Sliding Window Attention (R-SWA) keeps the KV cache flat, so memory and latency stay constant as output grows. It scores 93.23 on OmniDocBench v1.5 — beating DeepSeek-OCR by 6.22 points — at 12.7% higher throughput. This is the builder\'s read: what R-SWA actually does, how to run it locally with Transformers and vLLM, where it beats a cloud OCR API, when to skip it, and exactly how I\'d wire it into a production RAG ingestion pipeline.
Read the full version with code samples, diagrams, and architecture details: Baidu Unlimited-OCR: The Open-Source Model That Reads 40+ Page Documents in One Pass (2026)
More engineering notes: rohitraj.tech/en/notes
Top comments (0)