Developers Seek Self-Hosted OCR Stack Better Than Azure

Hot take: the market signal here is less “which OCR model wins” and more “which OCR stack is good enough to replace a paid API without blowing up latency.”

• –The complaint is about throughput as much as accuracy; a slower model that is only marginally better than Azure is not a real win.
• –Public benchmarks are finally moving: recent HF leaderboards and benches show newer open/document models like PaddleOCR-VL, dots.ocr, MinerU2.5, and olmOCR are now in the conversation, though results vary by task and corpus (https://huggingface.co/datasets/allenai/olmOCR-bench, https://huggingface.co/datasets/alphaXiv/2510.14528v1-ocr-dataset).
• –If the goal is self-hosting, the practical shortlist is likely a specialized OCR model or document parser rather than a general VLM.
• –The thread is also a reminder that OCR is benchmark-fragmented: text fidelity, table structure, reading order, and latency are often optimized separately.