Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With Built-In Speculative Decoding for Low-Budget GPUs

Jina AI (Elastic) introduced jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, and invoices to Markdown in one pass. The 3.4B parameter model (about 570M active per token) is based on DeepSeek-OCR and runs on budget GPUs like Nvidia L4. Weights are open under CC BY-NC 4.0; commercial use requires approval.

• 3.4B parameters, ~570M active per token, based on DeepSeek-OCR
• 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench
• 2.57 pages/s on a single A100 — best among 14 systems
• Weights ~6.8 GB in BF16, CC BY-NC 4.0 license, runs via Transformers or vLLM