HunyuanOCR

HunyuanOCR is a lightweight, 1B-parameter vision-language model from Tencent's Hunyuan Vision Team dedicated to OCR: text spotting, document parsing, information extraction, text-image VQA, and text-image translation. It pairs a native vision transformer with a small language model through an MLP adapter, eliminating the pre-processing modules of classic OCR pipelines. The December 2025 technical report claims it outperformed larger models such as Qwen3-VL-4B and took first place in the ICDAR 2025 DIMT Challenge small-model track, with SOTA OCRBench results among sub-3B VLMs.

The code is on GitHub under Tencent-Hunyuan and weights are on Hugging Face (tencent/HunyuanOCR), released under the Tencent Hunyuan Community License, which carries use conditions that commercial teams should review. A follow-up, HunyuanOCR-1.5, arrived in July 2026 with inference speedups and broader document understanding. License verify-card advised.