Sakhanda Wire
NVDA $153.61 -1.62% MSFT $506.27 -2.26% GOOGL $280.41 -1.17% META $623.08 -0.53% AMZN $257.66 -0.03%
← Back to the news

Jina AI releases a 3.4B-parameter document parser for low-cost GPUs

Jina AI, part of Elastic, has released jina-ocr-v1, a visual document parser that converts PDFs, scans, tables, charts and invoices into Markdown in one pass. The mixture-of-experts model contains 3.4B parameters, with about 570M decoder parameters active for each token. It uses DeepSeek-OCR as its base and includes a built-in FastMTP speculative-decoding head.

The system reports scores of 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench. It does not lead either benchmark, but its post-training improves the DeepSeek-OCR backbone by 7.4 points on olmOCR-Bench. On one A100 40 GB at concurrency 32, it processes 2.57 pages per second, the highest result among 14 systems measured by Jina AI.

Open weights are approximately 6.8 GB in BF16 and can run with Transformers or vLLM. The model is available for research and non-commercial use under CC BY-NC 4.0; commercial use requires contacting Jina AI. FastMTP requires vLLM 0.21 or later, while the Transformers implementation does not use the draft weights. Jina Reader and an OpenAI-compatible endpoint provide hosted access.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news