Jina AI released jina-ocr-v1 — an open document parsing model with 3.4 billion parameters, built on the DeepSeek-OCR architecture. The company's own contribution — FastMTP speculative decoding and GRPO post-training with component-wise verifiable rewards: this yielded a 7.4-point gain on olmOCR-Bench, first place in throughput among 14 systems (2.57 pages per second on a single A100), and nearly a 2x decoder speedup on budget GPUs. Weights are already published on Hugging Face, but the CC BY-NC 4.0 license only allows non-commercial use.



What happened
Jina AI released jina-ocr-v1 — a document parsing model for English and Chinese with 3.4 billion parameters, of which approximately 570 million are active in the decoder. The model accepts scans, photos, and PDFs and returns Markdown, where tables are encoded in HTML and formulas — in LaTeX. Compared to the base DeepSeek-OCR, it improved the result on olmOCR-Bench to 83.4 vs 76.0 and showed 91.14 on OmniDocBench v1.6. On a single A100 accelerator with 32 parallel requests, the model processed 2.57 pages per second, taking first place in throughput among 14 compared systems. Weights under the name jinaai/jina-ocr-v1 are published on Hugging Face.
Context
The architecture is inherited from DeepSeek-OCR and is not itself a Jina contribution: the visual encoder DeepEncoder with approximately 380 million parameters (SAM with 80 million and CLIP-L with 300 million) compresses a 1024×1024 page into 256 visual tokens, and the MoE decoder DeepSeek-3B converts them into text. The novelty of the release is concentrated in two elements. First — FastMTP: one shared dense draft block recursively predicts K=3 tokens ahead, and each guess is verified without quality loss; on an NVIDIA L4 accelerator this gives a speedup of about 1.95x without CUDA graphs and 1.17x with them. Second — post-training with the GRPO method with dense verifiable rewards: the final reward is decomposed into components like edit distance, TEDS metric for tables, formula matching, and unit tests, so the model is trained on automatically verifiable signals. For completeness: on OmniDocBench v1.6, the compact PaddleOCR-VL-1.6 scores 96.34, and on olmOCR-Bench chandra-ocr-2 shows 85.8 and dots.mocr — 83.9, meaning jina-ocr-v1 is not the absolute leader in accuracy. The developers themselves cite these losing figures, which speaks to accurate reporting.
Why this matters for the industry
The main industry signal — the economics of speculative decoding in OCR tasks: one reusable dense draft block costs almost nothing but gives a practically 2x decoder speedup without quality loss. For teams mass-converting archival documents to Markdown, this is noticeably cheaper: record throughput on a single A100 means fewer GPU-hours per page, and document ingest for RAG and agents turns into a cheap pipeline stage, where a scan, photo, or PDF at the input yields ready Markdown at the output. The second signal — reward decomposition proved more important than model size: a 3.4-billion model with GRPO outperformed olmOCR-2 with 8 billion parameters (82.4) on olmOCR-Bench. As a result, choosing an OCR model turns into a task of balancing speed, quality, price, and license, and the trick with a reusable draft block will likely be copied in other structured output models, intensifying competition for pages per second on budget GPUs.
Why this matters for users
You can use it in two ways. First — zero-infra: documents are processed through r.jina.ai or Jina Reader with an API key and the X-Respond-With: jina-ocr-v1 header, nothing needs to be installed. Second — self-hosted: the jinaai/jina-ocr-v1 weights are downloaded from Hugging Face and run on budget GPUs of the L4 class or consumer segment via vLLM 0.21+, where FastMTP acceleration is enabled. In addition to converting pages to Markdown, the model can caption images, answer questions about a document, and extract key fields. Key caveat for practitioners: the CC BY-NC 4.0 license prohibits commercial use of the open weights, so for a commercial product either API access through Jina or a model with a different license will be suitable.
What is still unknown / limitations
The model is trained on English and Chinese; support for other languages does not follow from the provided materials. The stated 2.57 pages per second is a throughput metric with batching, not single-request latency, and this distinction is important when evaluating claims. Open weights are licensed under CC BY-NC 4.0, i.e., only for non-commercial use. Predictions about other teams copying the FastMTP pattern are interpretation, not an established fact; the model's behavior on real archives with non-standard layouts and complex documents remains outside the provided benchmarks.
Sources
- jina-ocr-v1: Faster Document Parsing on Low-Budget GPUs (Jina AI blog post)
- jinaai/jina-ocr-v1 — model card and weights (Hugging Face)
- Jina-OCR-v1: Efficient Document Parsing with Speculative Decoding and Dense Verifiable Rewards (arXiv article)
Author
Look at AI, editorial team
