Researchers have introduced Jina-OCR-v1, a new document parsing model designed for efficient operation on low-budget GPUs. This model integrates the DeepSeek-OCR architecture with a FastMTP speculative decoding head, which recursively shares draft blocks to accelerate prediction. Jina-OCR-v1 achieves high scores on benchmarks like OmniDocBench v1.6 and olmOCR-Bench, and demonstrates a significant speed increase over traditional decoding methods on hardware such as the NVIDIA L4. The model's training involved instruction alignment, robustness fine-tuning, and a novel GRPO method using dense verifiable rewards for deterministic checks. AI
IMPACT This model's efficiency on low-budget GPUs could democratize advanced document parsing capabilities.
RANK_REASON The cluster describes a new model release and benchmark results published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
