Tencent open-sources Youtu-Parsing-Omni, a 5B docs-to-JSON parser
It tops OmniDocBench at 96.96 and self-hosts on one GPU; Baidu's MIT OCR 2 one-shots 3,600 pages, and open parsers are undercutting paid APIs.
Copy markdown
One model, every modality, one JSON
Tencent's Youtu-Parsing-Omni (5B, open weights) parses document pages, charts, geometry, audio and video into a single structured-JSON envelope — 0–1000 coordinates, HH:MM:SS timestamps. It posts a state-of-the-art 96.96 on OmniDocBench and 75.08 on OmniParsingBench, second only to Gemini 3 Pro, and serves an OpenAI-compatible endpoint via vLLM. Mind the custom "youtu-parsing" license (no EU use).
Baidu's OCR 2 one-shots a 3,600-page archive
Unlimited OCR 2 is MIT-licensed and parses a document of any length in a single pass — Baidu's demo ingested all seven Harry Potter books (3,600+ pages) at once. Its SR-SWA sparse attention decodes ~7x faster than full attention at 100 pages, and v1 already cleared 10M Hugging Face downloads. For ingestion pipelines: no chunking, no per-page bill, and real free commercial use.
Hosted parsing APIs lose the moat
On OmniDocBench, sub-2B open specialists — PaddleOCR-VL (96.34), MinerU 2.5 (95.75), TeleOCR (96.91) — now sit within a point of the new leaders and comfortably beat Gemini 3 Pro (92.91) and GPT-5.2 (86.59) at page OCR. If your ingestion layer calls Google Document AI or Cohere Parse per page, that line item is now replaceable with one GPU.
OpenAI: up to $500/mo in free API credits
Per OpenAI's developer changelog (Oct 28), eligible ChatGPT Pro and Business subscribers can link an API organization they own and claim up to $500 each month in free API credits. If you already pay for Pro, that's a standing monthly budget for agents and batch jobs you may not be tapping.