Jina AI Releases jina-ocr-v1 With Highest Page Throughput Among 14 Models

• Jina AI released jina-ocr-v1, a document parsing model built on DeepSeek-OCR that achieves 2.57 pages per second on an NVIDIA A100 GPU.
• The model utilizes FastMTP speculative decoding and a Mixture of Experts architecture to deliver lossless, byte-identical output while reducing computational load.
• In Jina's internal benchmarks, jina-ocr-v1 ranked first in page throughput among 14 tested systems, though it ranks third in accuracy benchmarks.
• The release aligns with Elastic's strategy to lower per-page costs for document ingestion in retrieval-augmented generation (RAG) pipelines.
• The model is available under a CC BY-NC 4.0 license, with commercial usage requiring contact with Jina AI.

Jina AI has officially released jina-ocr-v1, a document parsing model designed to convert PDFs, scanned documents, tables, and charts directly into Markdown. The model is post-trained on DeepSeek-OCR and retains its Mixture of Experts (MoE) architecture. This structure comprises approximately 3.4 billion total parameters, with only 570 million active per token. This design allows the model to operate at the running cost of a small model while retaining large capacity.

A key differentiator is the integration of FastMTP speculative decoding. This method drafts three tokens ahead and verifies them greedily. The mechanism ensures the output is lossless and byte-identical to standard autoregressive decoding. It averages 2.73 tokens committed per step at K=3, significantly accelerating inference.

In internal benchmarks, jina-ocr-v1 scored 91.14 on OmniDocBench v1.6. It also scored 83.4 on olmOCR-Bench, a 7.4-point improvement over the base DeepSeek-OCR model. The model outperformed 14 other tested systems in throughput tests. On a single NVIDIA A100 GPU with 32 concurrent requests, it processes 2.57 pages per second. This is approximately 22% faster than DeepSeek-OCR's 2.10 pages per second.

How Does the Architecture Reduce Inference Costs?

The model's speed advantage comes from output concision and reduced decoding passes. It emits only about 1,085 tokens per page compared to competitors like Surya OCR 2, which emit around 3,568. The architecture uses a DeepEncoder with a convolutional compressor and CLIP-L stage to reduce visual token count. This allows for dynamic resolution handling.

The DeepEncoder, containing approximately 380 million parameters, chains SAM, a 16x convolutional compressor, and CLIP-L. It compresses visual input into 256 tokens per page. Dynamic resolution adds up to 1,156 tokens for complex pages. Post-training involves instruction alignment and robustness fine-tuning on degraded pages using GRPO. Rewards are deterministic and code-scored against reference transcriptions.

On cheaper hardware like the NVIDIA L4, speculative decoding provides a 1.95x speedup in eager mode. The model is best suited for bulk document ingestion and pipelines where cost-per-page is critical. It supports 100 languages and handles complex structures like formulas and tables with high accuracy.

What Are the Accuracy Trade-Offs and Limitations?

While throughput is strong, accuracy is competitive rather than leading. It scores 91.14 on OmniDocBench v1.6, behind PaddleOCR-VL-1.6 at 96.34. On olmOCR-Bench, it trails chandra-ocr-2, which scores 85.8. The model struggles with degraded historical scans, scoring only 42.6 on the OldScans subset.

The benchmarks were self-conducted with uncontrolled hardware settings. This suggests the 'first place' speed ranking is specific to Jina's infrastructure. The model drops headers and footers and may require human review for heavily degraded scans. It is available via Jina Reader, an OpenAI-compatible API, and self-hosted weights under CC BY-NC 4.0.

For investors, the key value is not technical disruption but cost reduction. Elastic acquired Jina to lower the per-page cost of document ingestion in its RAG pipelines. The improvement is incremental to Elastic's unit economics rather than a competitive moat. The underlying architecture is open-source and easily replicable. Weights are available on Hugging Face, requiring vLLM 0.21+ for self-hosting.