HyOCR-1.5 - Tencent Hunyuan's open-source lightweight end-to-end OCR expert model - AiBoss

What is HyOCR-1.5?

HyOCR-1.5 is a lightweight, end-to-end OCR expert model open-sourced by Tencent Hunyuan. With only 1 B parameters, it achieves full-stack capabilities including document parsing, text recognition, information extraction, image translation, chart parsing, ancient character recognition, video subtitle extraction, and multi-page document question answering. Based on DFlash speculative decoding, the model achieves up to 6.37 times faster inference speed for long structured outputs; it ranks first in end-to-end OmniDocBench v1.6 with a score of 94.74.

Main functions of HyOCR-1.5

Document parsingConvert dense documents, multi-column layouts, tables, and formulas into structured text (Markdown/HTML/LaTeX) with one click.
Text detection and recognitionOutputs all text content and corresponding coordinates in an image, covering scenarios such as documents, street views, handwriting, advertisements, games, and videos.
Information ExtractionExtract specified fields from invoices, certificates, and receipts and return them in JSON format.
Photo translationExtract text from images and translate it into multiple languages, preserving document layout and formula formatting.
Chart Analysis: Parse flowcharts and statistical charts into Mermaid or Markdown format.
Ancient Chinese character recognitionSupports recognition of seven styles of Chinese characters (Oracle Bone Script, Bronze Script, Seal Script, Clerical Script, Regular Script, Running Script, and Cursive Script).
Video subtitle extraction: Accurately extract subtitle text from video frames.
Multi-page document Q&ACross-page retrieval, comparison, and evidence aggregation question answering based on multi-page PDFs.

The technical principle of HyOCR-1.5

Lightweight end-to-end architectureIt adopts the compact design validated by HyOCR-1.0, consisting of the native resolution visual encoder Hunyuan-ViT, an adaptive MLP connector, and the lightweight language model Hunyuan-0.5B. It directly maps multimodal inputs to structured outputs such as Markdown, HTML, and LaTeX without requiring any task-level post-processing modules.
4K native resolution visual encodingThe visual encoder is built on Hunyuan-ViT, with the maximum input resolution expanded from 2K to 4K, while maintaining the original aspect ratio and spatial layout, enabling the model to capture fine-grained structural details of high-density documents, ultra-large tables, and complex layouts.
DFlash Speculative DecodingThe block-diffusion draft model with approximately 90.7M parameters is introduced, which can predict the entire candidate token in a single parallel forward pass. The target model then verifies and accepts the longest correct prefix in a single pass. While strictly maintaining the output distribution of the target model, the decoding latency of the long structured OCR is significantly reduced.
Agentic Data FlowThe intelligent agent drives a closed loop of data production, transforming model shortcomings into actionable data requirements. It autonomously completes material search, tool-assisted cleaning, difficult example mining, and data pipeline development, and continuously iterates with algorithm engineers.
Three-stage training formulaThe pre-training stage re-plans Stage 3, injecting new capability data and expanding to 4K resolution and 128K context; the SFT stage thoroughly cleans the data and unifies the prompt interface; the RL stage adopts IcePop (GRPO style) optimization, improving output quality through three complementary rewards: factuality, consistency evaluation, and degradation suppression.

How to use HyOCR-1.5

Environmental preparationInstall Python 3.12+, CUDA 12.9, PyTorch 2.7.1 and vLLM (≥0.12.0).
Model Startup:implement vllm serve tencent/HunyuanOCR Start the service or load the Hugging Face weights.
Image inputUse PIL to read images and build a message conversation structure containing images and text.
Task InstructionsSelect the prompt based on the scenario, such as document parsing, text detection, information extraction, translation, etc.
Invoking inference: Generate results using vLLM or Transformers, setting temperature=0 and max_tokens=16384.
Results of cleaningUse the built-in deduplication function to clean up possible duplicate substrings and obtain the final structured output.
Local deploymentRuns on CPU or consumer-grade GPU via llama.cpp, no server required.

HyOCR-1.5's core advantages

Speed ReasoningDFlash speculative decoding speeds up long document generation by 6.37x in Transformers and 2.14x in vLLM, with an end-to-end page generation time of just 1.4 seconds.
Lightweight and deployableThe 1B parameter can be run locally on CPUs, consumer-grade graphics cards, and laptops via llama.cpp.
SOTA accuracyOmniDocBench v1.6 ranked first in end-to-end performance (94.74), with a table TEDS score of 93.67, showing outstanding performance in complex tables and reading order.
Full-stack open sourceThe training formula, inference framework, and model weights are fully disclosed, and are reproducible, fine-tunable, and scalable.
Long tail capabilityAgentic Data Flow adds rare capabilities such as 331 low-resource languages, ancient scripts, and multi-image question answering.
High-resolution long contextSupports 4K image resolution and 128K context, adapting to high-density long documents.
Hallucination inhibitionThe CHAOS-Bench model achieves a page-average recall of 14.15, which is better than existing models and the output is more faithful to what is seen.

HyOCR-1.5 project address

GitHub repositoryhttps://github.com/Tencent-Hunyuan/HunyuanOCR
HuggingFace model libraryhttps://huggingface.co/tencent/HunyuanOCR
arXiv technical paper: https://arxiv.org/pdf/2607.04884

HyOCR-1.5 Comparison with Similar Products

Dimension	HyOCR-1.5	DeepSeek-OCR-2
Parameter size	1B (Lightweight)	3B (3 times larger)
OmniDocBench	94.74	87.01
End-to-end delay	1.408s / page	5.460s/page (3.9 times slower)
Table Analysis TEDS	93.67	Approximately 84.97
Inference acceleration	DFlash speculative decoding, Transformers 6.37× speedup	No dedicated acceleration, pure autoregressive decoding
Deployment costs	Can run on CPU / laptop	Server-grade GPU required

Application scenarios of HyOCR-1.5

Intensive document digitizationConvert high-density, multi-column layouts such as contracts, papers, and reports into Markdown/HTML/LaTeX structured text with a single click, preserving the reading order and layout.
Analysis of complex tables and formulasIt accurately reproduces large tables and mathematical formulas in financial statements and academic papers, and supports HTML table and LaTeX formula output.
Multilingual cross-border document processingIt covers 331 languages, automatically recognizes and parses mixed language documents, and is suitable for global business scenarios such as foreign trade, legal affairs, and publishing.
Ancient script research and protectionIt can identify seven styles of Chinese characters, including oracle bone script, bronze script, seal script, and clerical script, to assist museums and archaeological institutions in digital archiving and research of historical documents.