Tencent Hunyuan Open-Sources 1B End-to-End OCR

腾讯混元 · wechat · 2026-07-13

Tencent Hunyuan released HyOCR-1.5, a 1B-parameter end-to-end OCR expert model with a fully open-source stack covering training, inference, and weights. It supports over 8 text-centric tasks and leverages DFlash-based speculative decoding to achieve up to a 6.37× inference speedup. It also scores 94.74 on OmniDocBench v1.6, ranking among the top tier of end-to-end OCR models.

The article highlights three key aspects:

Faster: Uses a lightweight draft model with verification-based speculative decoding to accelerate autoregressive generation for long documents, delivering significant gains for long outputs and table pages.
Stronger: Introduces AgenticDataFlow, enabling agents to automatically collect materials, develop data pipelines, and iterate based on model weaknesses, thereby improving capabilities in low-resource languages, ancient texts, and multi-image QA.
More Comprehensive: Integrates 4K resolution, 128K context, and RL on the training side; evaluations cover document parsing, ancient texts, charts, multilingual scenarios, multi-image QA, and hallucusion suppression.

The article also shares multiple results: achieving SOTA in ancient text recognition at the 1B scale on Chronicles-OCR; performing close to or even exceeding some 8B models on ChartArena; and reaching a 99.8% accuracy in text-free image processing.

Original post →

Tencent Hunyuan Open-Sources 1B End-to-End OCR

More from Multimodal