Tencent's HunyuanOCR Beats Gemini-2.5-Pro at Document Parsing With 1B Parameters
Tencent's 1B-parameter HunyuanOCR beats models 200x its size on document parsing, spotting, and translation — and it's fully open-source
Takeaways
- 1B-parameter OCR model from Tencent Hunyuan beats Gemini-2.5-Pro and Qwen3-VL-235B on document parsing and information extraction benchmarks.
- End-to-end single model covers text spotting, document parsing, info extraction, subtitle extraction, and photo translation via prompt switching.
- 94.10 on OmniDocBench (document parsing), 92.87 on video subtitle extraction, and 70.92% on multi-scene text spotting — all SOTA.
- First OCR model trained with GRPO reinforcement learning, using verifiable rewards to enforce structured output formatting.
- Fully open-source: weights on Hugging Face, code on GitHub, live demo available — free to use.
- Requires 20GB VRAM under vLLM; Transformers backend has a known accuracy gap vs. vLLM that is still being fixed.
HunyuanOCR is Tencent Hunyuan's open-source OCR model that punches well above its weight class. At just 1 billion parameters, it tops leaderboards across document parsing, text spotting, information extraction, and image-to-text translation , outperforming models like Qwen3-VL-235B and Gemini-2.5-Pro on several key benchmarks, while fitting in a fraction of the memory footprint.
The model is available now on Hugging Face, with weights, inference code, and a live demo all publicly accessible. It runs on a single GPU with 20GB of VRAM under vLLM, and the model weights themselves only take up about 6GB of disk space. There is no usage cost , it is fully open-source.
One model, every OCR task
The core pitch of HunyuanOCR is consolidation. Traditional OCR pipelines are a chain of specialized modules: a layout detector, a text detector, a recognizer, a post-processor. Each step introduces its own errors, and the mistakes compound. HunyuanOCR adopts a pure end-to-end paradigm that eliminates dependencies on pre-processing modules like layout analysis, fundamentally resolving error propagation common in traditional pipelines and simplifying system deployment.
HunyuanOCR is a 1 billion parameter vision-language model designed specifically for optical character recognition and document understanding, performing spotting, parsing, information extraction, visual question answering, and text-image translation in a single end-to-end pipeline. You switch tasks by changing the prompt , no model swapping, no separate inference calls.
The tasks it covers out of the box:
- Text spotting: Detects and localizes every text region in an image, outputting bounding box coordinates alongside the recognized text.
- Document parsing: Converts scanned or photographed documents into structured Markdown, with formulas in LaTeX and tables in HTML.
- Information extraction: Pulls structured fields (names, amounts, dates) from receipts, cards, and invoices into JSON.
- Video subtitle extraction: Reads subtitles directly from video frames, including bilingual ones.
- Photo translation: Reads text in an image and translates it end-to-end, supporting 14 languages into Chinese or English.