Jina-OCR-v1 Achieves 2.57 Pages Per Second via Speculative Decoding
September 4, 2026
Jina-OCR-v1 uses a 3B MoE decoder with a FastMTP speculative decoding head to achieve high throughput on low-budget GPUs like the NVIDIA L4. The model scores 91.14 on OmniDocBench v1.6 and employs GRPO with dense verifiable rewards for structural and formulaic accuracy.
HOW THIS AFFECTS YOU
●
builderYou can deploy high-throughput document parsing on inexpensive hardware with minimal loss in structural accuracy.