jinaai/jina-ocr-v1

Jina releases jina-ocr-v1, a document parsing model built on DeepSeek-OCR with speculative decoding. It reports 91.14 on OmniDocBench v1.6 and 2.57 pages/s throughput on A100. Weights and custom loading code are available via Hugging Face, supporting Transformers and vLLM backends for local or hosted OCR tasks.