Tencent Open-Sources HunyuanOCR, A Light-weight Business-Grade Imaginative and prescient Mannequin

SHENZHEN – Tencent’s Hunyuan massive mannequin staff has formally launched and open-sourced HunyuanOCR, a specialised light-weight vision-language mannequin for optical character recognition (OCR) containing simply 1 billion parameters.

The mannequin combines a local Imaginative and prescient Transformer (ViT) structure with a light-weight massive language mannequin (LLM), delivering commercial-level efficiency in textual content detection, doc parsing, and data extraction. It lately received first place within the small mannequin observe of the ICDAR 2025 DIMT problem and achieved state-of-the-art outcomes on the OCRBench benchmark for fashions underneath 3B parameters.

HunyuanOCR introduces three key breakthroughs:

1. Unified multitasking functionality – supporting textual content detection, complicated format evaluation, open-field data extraction, and picture translation inside a single environment friendly framework
2. Finish-to-end structure – eliminating conventional preprocessing pipelines and lowering error accumulation
3. Reinforcement studying optimization – demonstrating that RL can considerably improve efficiency throughout a number of OCR duties

The mannequin has gained speedy neighborhood traction, rating among the many prime 4 trending fashions on Hugging Face and receiving over 700 stars on GitHub inside a brief interval. It has additionally been built-in into the vLLM inference engine.

Accessible now on Hugging Face and ModelScope, HunyuanOCR offers researchers and builders with a strong, deployable OCR resolution that balances excessive accuracy with computational effectivity – significantly helpful for edge deployment and industrial functions.