Recently, the lightweight end-to-end OCR expert model HyOCR-1.5 was officially released. Through a series of technological innovations, it achieves a significant improvement in performance and efficiency while maintaining a lightweight architecture.

As the first full-stack open-source model in this field, HyOCR-1.5 not only releases model weights but also fully opens up training recipes, data construction methods, and inference acceleration frameworks to the community. This move greatly reduces the development threshold, allowing developers to easily reproduce, fine-tune, or even deploy it on consumer-grade GPUs or regular laptops.

image.png

To address the latency bottleneck caused by long autoregressive decoding, the research team introduced a speculative decoding framework called "DFlash." By using a lightweight draft model with about 90.7M parameters for parallel prediction, DFlash achieves a multiple-fold increase in inference speed while ensuring output accuracy. In the authoritative evaluation OmniDocBench, this technology brings a 6.37x acceleration under the Transformers architecture, making it a leader among end-to-end OCR models.

In terms of model capability evolution, HyOCR-1.5 adopts an innovative strategy called "agent-driven data flow." The research team transforms the model's weak points into specific task objectives, which are then autonomously decomposed, collected, and verified by agents. This closed-loop training mode successfully addresses the shortcomings of long-tail scenarios such as ancient text recognition, low-resource language processing, and cross-page multi-image QA. Combined with training optimization for 4K resolution input and a 128K context window, the model's robustness in handling complex documents has been significantly enhanced.

Evaluation data shows that HyOCR-1.5, with only 1B parameter scale, demonstrates "superior" performance in multiple tasks. On the OmniDocBench v1.6, it not only remains in the top tier of end-to-end models, but its performance in ancient text recognition and chart parsing tasks can even rival general models with 8B scale.