Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With Built-In Speculative Decoding for Low-Budget GPUs
Jina AI open-sourced jina-ocr-v1, a 3.4B MoE OCR model with speculative decoding that parses documents to Markdown at 2.57 pages/second on an A100.
Jina AI released jina-ocr-v1, a 3.4B-parameter mixture-of-experts visual document parser post-trained from DeepSeek-OCR, with about 570M decoder parameters active per token. It converts PDFs, scans, tables, charts and invoices to Markdown in one pass, scoring 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench (7.4 points above its DeepSeek-OCR backbone). A built-in FastMTP speculative decoding head drafts three tokens per step with greedy verification, yielding a lossless 1.95x speedup on an NVIDIA L4 and 2.57 pages/second on an A100 40GB, the highest of 14 measured systems. Open weights (~6.8GB BF16) ship under CC BY-NC 4.0 for research and non-commercial use, runnable on Transformers and vLLM 0.21+.
- 3.4B total parameters with ~570M active per token, post-trained from DeepSeek-OCR
- FastMTP speculative decoding drafts 3 tokens per step; greedy verification keeps output lossless
- Scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench
- Throughput of 2.57 pages/second on one A100, highest of 14 measured systems
- Weights (~6.8GB BF16) released under CC BY-NC 4.0; commercial use requires a license
- AI models
- jina-ocr-v1DeepSeek-OCR
From the original article
4 min read at marktechpost.com
Jina AI has released jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, charts and invoices into Markdown. The model has 3. 4B total parameters, with about 570M active per token, and builds on DeepSeek-OCR.
By Asif Razzaq · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.