Jina AI publishes a 3.4B OCR model that bars commercial use

single source· 1 articles · confidence: high · first seen 2026-09-18 21:21 UTC

What this means for you

Check the licence before you build on it: CC BY-NC 4.0 bars commercial use of the weights, and Jina Reader is the hosted route instead. Throughput is reported at 2.57 pages per second on a single A100, so plan GPU capacity for anything large. Both scores come without evaluation dates.

Jina AI has released jina-ocr-v1, a model based on DeepSeek-OCR that turns PDFs, scans, tables, charts and invoices into Markdown. It has 3.4B total parameters but runs about 570M per token — a mixture-of-experts design, where only part of the network fires on each request. A built-in speculative decoding head drafts three tokens at a time and checks them, so the output is unchanged. Jina reports 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, without evaluation dates, and 2.57 pages per second on one A100. Weights are on Hugging Face under CC BY-NC 4.0, which bars commercial use; hosted access is via Jina Reader.

Key facts

  • ·jina-ocr-v1 has 3.4B total parameters and about 570M active per token. source
  • ·It converts PDFs, scans, tables, charts and invoices into Markdown. source
  • ·Jina reports 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench. source
  • ·It parses 2.57 pages per second on a single A100. source
  • ·Weights are on Hugging Face under CC BY-NC 4.0, with hosted access through Jina Reader. source
  • ·It builds on DeepSeek-OCR and uses a FastMTP speculative decoding head drafting three tokens per step. source

What the sources say

  • MarkTechPost — Model card detail: size, licence terms, throughput on one A100 and two parsing scores.

Sources

The original reporting. Follow these — they did the work.

← the wire