Jina AI Releases jina-ocr-v1: A 3.4B MoE Document Parser With Built-In Speculative Decoding for Low-Budget GPUs
3.4B parameters, 570M active per token: Jina's new OCR model hits 91+ on benchmarks while running on a single A100.
Why it matters
Open-weight document parsing model with MoE and speculative decoding brings frontier OCR capability to practitioners on modest hardware. Benchmark scores and efficiency metrics matter for teams evaluating document-to-text pipelines.
The key facts
7 to knowModel: jina-ocr-v1, 3.4B total parameters with ~570M active per token (MoE)
Architecture: built on DeepSeek-OCR with FastMTP speculative decoding (3 tokens per step, lossless)
Benchmarks: 91.14 on OmniDocBench v1.6, 83.4 on olmOCR-Bench
Throughput: 2.57 pages/sec on 1x A100 GPU
Weights: Hugging Face, CC BY-NC 4.0 license
Access: hosted via Jina Reader API
Target: low-budget GPU inference
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: 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. A built-in FastMTP speculative decoding head drafts 3 tokens per step…