Jina-OCR-v1 (arXiv 2609.03181, Fri 4 Sep digest; Jina AI, co-author Han Xiao) is a document-parsing model combining a DeepSeek-OCR-style compressed vision encoder with a 3B MoE decoder (~0.57B parameters active per token) and a FastMTP speculative-decoding head that recursively shares a single draft block across K=3 prediction steps; greedy verification keeps decoding lossless. Post-training runs instruction alignment, robustness fine-tuning, and GRPO with dense verifiable rewards — deterministic formula/table/structure checks that award partial credit. Scores: 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench at 2.57 pages/s (the highest throughput in its comparison); FastMTP doubles decode speed on an NVIDIA L4 GPU. The models are public on Hugging Face.
This is a complete, deployable production recipe for a workload almost every RAG pipeline depends on: document parsing at 2.5+ pages/s on a single L4-class GPU, with the full post-training stack (MoE + MTP + dense-reward GRPO) documented and weights public. For teams currently paying OCR API bills, this is a self-host alternative with benchmark numbers to check against.
| Arxiv | 2609.03181, announced in the Fri 4 Sep 2026 digest |
|---|---|
| Architecture | compressed vision encoder (DeepSeek-OCR style) + 3B MoE decoder (~0.57B active/token) + FastMTP speculative decoding head (one draft block reused recursively across K=3 steps, greedy lossless verification) |
| Post Training | instruction alignment + robustness fine-tuning + GRPO with dense verifiable rewards (deterministic formula/table/structure checks, partial credit) |
| Results | OmniDocBench v1.6 91.14; olmOCR-Bench 83.4; 2.57 pages/s (highest in its comparison); FastMTP doubles decoding on an NVIDIA L4 |
| Availability | models public under the jinaai HF org |