What is HunyuanOCR 1.5?
HunyuanOCR 1.5 is a 1.12-billion-parameter OCR vision-language model from Tencent's Hunyuan team, released on 7 July 2026 with a technical report on arXiv. It keeps the architecture of HunyuanOCR 1.0, a vision encoder feeding a Hunyuan-0.5B language model, and changes the training: images up to 4K resolution, a 128K-token context, and data built by an agent-driven pipeline aimed at ancient scripts, low-resource languages and questions across several images.
One model handles document parsing to Markdown with HTML tables and LaTeX formulas, text spotting with coordinates, field extraction, chart parsing and translation of text in images. A DFlash draft model, shipped in the same repository, drafts several tokens at once for the main model to verify, which speeds up long outputs without changing them.
The weights are under the Tencent Hunyuan Community License. It does not apply in the European Union, the United Kingdom or South Korea, and services with more than 100 million monthly active users need a separate licence from Tencent.
Is HunyuanOCR 1.5 worth it?
HunyuanOCR 1.5's appeal is breadth at a small size. In Tencent's report it scores 94.74 on OmniDocBench v1.6, the top figure among the end-to-end OCR models in its table, ahead of Unlimited-OCR at 93.92, Qianfan-OCR at 93.90 and FireRed-OCR at 93.26. That table leaves out two-stage pipelines; PaddleOCR-VL-1.6 reports 96.33 on the same benchmark, and Jina's own comparison table lists the two at 94.74 and 96.34. It also keeps HunyuanOCR 1.0's general skills, scoring 861 on OCRBench against 860 before.
The speed claim is narrower than the abstract suggests. DFlash gives 2.14 times the plain decoding speed under vLLM and 6.37 times under Transformers, which starts from a slow 0.029 pages a second. In Tencent's single-request test on an H20, 0.706 pages a second puts it ahead of GLM-OCR at 0.604 and PaddleOCR-VL-1.6 at 0.562, but at eight concurrent requests it reaches only 1.797, a modest figure next to the batched throughput other teams publish, though those runs used different hardware.
All accuracy and speed numbers are Tencent's. The model is not yet on the independent BHL OCR board, and Tencent's own CHAOS-Bench, a hallucination test it introduced in this release, is one where it naturally leads.
The licence is the deciding factor for many readers. The Tencent Hunyuan Community License excludes the European Union, the United Kingdom and South Korea outright, including use of the model's output there, so a European company cannot deploy it at all. For teams elsewhere it is one of the few sub-2B models that also does text spotting, field extraction and image translation, which PaddleOCR-VL-1.6 and MiniCPM-OCR 1.0 do not cover in one model.
Where it earns its place
A localisation team at a Singapore games publisher
Translate text inside screenshots and marketing images
Serve the model under vLLM and use the trans_other2en or doc_trans_en2zh task types to return translated text directly. Singapore is inside the licence territory; a London office of the same company is not.
A back-office engineer at a Chinese insurer
Extract fields from cards and receipts into structured data
Use the structured_parse or information-extraction prompts on one GPU with the DFlash draft enabled. Tencent reports 92.40 on cards and 92.55 on receipts in its extraction tests, against 80.59 and 80.66 for Gemini 2.5 Pro.
A developer at a US research start-up working on laptops
Parse papers and charts offline without a server GPU
Convert the checkpoint to GGUF and run llama-server, or the DFlash-adapted llama.cpp fork, on a laptop. The weights are about 2.2 GB in BF16, and the licence costs nothing below 100 million monthly users.
Before you commit
- Setup
- Create a uv environment with vLLM 0.25.1+ and flash-attn 2.8.3 (CUDA 13), download tencent/HunyuanOCR excluding v1.0, then start serve.sh or serve_DFlash.sh from the GitHub repo.
- Integrations
- Transformers 5.13+ native class; vLLM OpenAI-compatible server with batch scripts; llama.cpp via GGUF conversion; verl-based reinforcement-learning stack and SFT code on GitHub.
- Data and privacy
- Tencent Hunyuan Community License: not valid in the EU, UK or South Korea, an acceptable-use policy applies, and services over 100 million monthly active users must ask Tencent. Self-hosted runs keep documents local.
- Cost at scale
- Tencent measures 0.706 pages a second with DFlash at one request and 1.797 at eight on one H20; no hosted price is published, so cost is your own GPU time.
- Limits
- The unified environment needs CUDA 13; the Transformers path is far slower than vLLM; English and Chinese are the main documented languages; the shipped client allows up to 32,768 output tokens on dense pages.
What’s good
- 94.74 on OmniDocBench v1.6, top end-to-end model in Tencent's table
- 1.12B parameters, with a 128K context and 4K input resolution
- Parsing, spotting, extraction, chart parsing and translation in one model
- DFlash draft model included; 2.14 times faster under vLLM
What’s not
- Licence excludes the EU, UK and South Korea
- Services over 100 million monthly users need Tencent's permission
- All benchmark and speed figures are self-reported
- Behind PaddleOCR-VL-1.6's two-stage pipeline on OmniDocBench v1.6
Best for
Teams outside Europe, the UK and South Korea who want one small model for document parsing, text spotting, extraction and image translation.
HunyuanOCR 1.5 alternatives
HunyuanOCR 1.5: common questions
- Is HunyuanOCR 1.5 worth it?
- HunyuanOCR 1.5 is the strongest end-to-end OCR model near 1B parameters on Tencent's own numbers, and legally off-limits to anyone in Europe, the UK or South Korea. We rate it 4.1 out of 5.
- How much does HunyuanOCR 1.5 cost?
- Free under the Tencent Hunyuan Community License outside the EU, UK and South Korea; services over 100 million monthly users need a licence from Tencent. No hosted API price published.
- What are the downsides of HunyuanOCR 1.5?
- Licence excludes the EU, UK and South Korea; services over 100 million monthly users need Tencent's permission; all benchmark and speed figures are self-reported; behind PaddleOCR-VL-1.6's two-stage pipeline on OmniDocBench v1.6.
- Who is HunyuanOCR 1.5 best for?
- Teams outside Europe, the UK and South Korea who want one small model for document parsing, text spotting, extraction and image translation.
- How much does HunyuanOCR 1.5 cost for a team or in production?
- Tencent measures 0.706 pages a second with DFlash at one request and 1.797 at eight on one H20; no hosted price is published, so cost is your own GPU time.
- Is HunyuanOCR 1.5 safe for company data?
- Tencent Hunyuan Community License: not valid in the EU, UK or South Korea, an acceptable-use policy applies, and services over 100 million monthly active users must ask Tencent. Self-hosted runs keep documents local.
- What are the best HunyuanOCR 1.5 alternatives?
- The closest alternatives we have reviewed are PaddleOCR-VL-1.6, MiniCPM-OCR 1.0, GLM-OCR, Qianfan-OCR and Unlimited-OCR 2.
Answers checked Nov 2, 2026.
Make HunyuanOCR 1.5? Show this rating on your site
