Tencent Hunyuan Releases HyOCR-1.5 End-to-End OCR Large Model Fully Open-Sourced

Tencent Hunyuan released HyOCR-1.5 this week. The 1B parameter model is the first end-to-end OCR large model to publish training code, inference code, and weights together.

The release also includes a new speculative decoding framework that cuts inference time under Transformers by 6.37 times. End-to-end page processing now averages 1.408 seconds.

Release Details and Technical Scope

HyOCR-1.5 covers more than eight text-centric tasks. It supports 4K resolution images and a 128K context window.

The model adds Agentic Data Flow to extend coverage to 331 languages, ancient scripts, and multi-image question answering. On OmniDocBench v1.6 the system scored 94.74, the highest end-to-end result reported so far.

Tencent Hunyuan states that every component needed to reproduce training and deployment is now public on its repository.

Performance Gains from DFlash Framework

The new DFlash speculative decoding framework runs on top of standard Transformers. It records a 6.37 times speedup compared with the baseline.

Under vLLM the same framework yields a 2.14 times gain. Both figures were measured on the same hardware and document distribution used for the OmniDocBench benchmark.

Tencent Hunyuan claims these numbers hold across typical office documents at 4K resolution. Independent reproduction will be required to confirm the gains on other datasets.

Industry Pressure on Closed OCR Providers

The open release directly challenges vendors that keep training pipelines and weights private. Several commercial OCR services still require paid API calls for complex layouts and low-resource languages.

Developers who previously relied on those services now have a documented 1B model they can fine-tune locally. The combination of public weights and the DFlash decoder lowers the barrier for on-premise deployment.

Larger closed providers face a choice between matching the openness or defending paid tiers with extra features.

Extension to Low-Resource and Ancient Scripts

Agentic Data Flow routes difficult samples through an agent loop that generates additional training pairs. Tencent Hunyuan reports this loop lifted performance on languages with fewer than 10,000 public samples.

The same mechanism supports ancient Chinese and Egyptian scripts by treating them as additional low-resource targets. Early tests show usable accuracy on scanned historical documents without new labeled data.

Whether the loop scales beyond the reported 331 languages remains untested by outside groups.

Remaining Questions on Verification and Adoption

The 94.74 OmniDocBench score comes from a single evaluation run published by the team. No third-party audit of the test split or scoring script has appeared yet.

The 6.37 times speedup figure also rests on internal measurements. Reproducible scripts are available, but community benchmarks on varied hardware have not been published.

Adoption speed will depend on how quickly downstream projects integrate the released code and confirm the claimed latency on their own document sets.

What to Watch in the Next Three Months

Watch for independent reproductions of the DFlash benchmarks on public cloud instances. Consistent results above 5 times would strengthen the performance claim.

Watch for fine-tuned versions targeting specific domains such as medical forms or historical archives. Early domain models would signal practical uptake.

Watch for any response from closed OCR vendors in pricing or feature updates. A sudden price cut or new open model from a competitor would indicate direct pressure from the HyOCR-1.5 release.