HunyuanOCR
Tencent · China ·
Commercial use: Conditional — commercial use with a trigger
Custom Tencent licence, tagged "other". Read the repository terms before commercial use — do not assume permissive because the download is open.
A model that reads scanned documents — invoices, forms, records — and turns them into structured text.
Why it should matter to you
Most businesses sitting on a paper or PDF archive assume digitising it is a bespoke project. Specialist OCR models have made the reading part close to solved. What has not changed is everything after: deciding that four spellings of a name are one customer.
How it connects to our work
Reading the page is now the easy half. On the patient registry extraction, OCR was days and identity resolution was months — and that ratio is typical rather than unusual.
From our field notesLegacy system access is the most overlooked pre-build blocker.
Advantages
- Purpose-built for document OCR rather than general vision, and it shows on dense scanned text
- 660K downloads, indicating real adoption for extraction work
- Handles multilingual and structured documents better than general VLMs
Disadvantages
- Custom licence needs reading
- Chinese-origin, which draws procurement questions in document work involving personal data
- Specialist: no use outside document understanding
Reach for it when
Extraction pipelines over scanned archives, where a general vision model wastes capacity and money.
Where it falls down
Anything but documents. And check the licence before it touches a commercial pipeline — this is exactly the case where "downloadable" is doing a lot of unearned work.
Running it
12 GB GPU. See the hardware sizing tables for how that maps to specific chips and cards, and the quantisation guide for what you give up at each bit width.
Jurisdiction: China
This is the crucial split most write-ups miss: using a Chinese lab's HOSTED API sends your data to Chinese infrastructure and is a real procurement question. Downloading their OPEN WEIGHTS and running them on your own hardware, or on a Western host, sends nothing anywhere. DeepSeek and Z.ai publish under MIT; Alibaba publishes most of Qwen3 under Apache 2.0. Those are among the most permissive licences on this page.
Watch for: Several US states and a number of government bodies restrict Chinese-hosted AI services on official devices. That restriction is about the hosted service, not about the weights running in your own VPC — but expect to have to explain the difference to a security reviewer.
Frequently asked
Can I use HunyuanOCR commercially?
Custom Tencent licence, tagged "other". Read the repository terms before commercial use — do not assume permissive because the download is open.
What hardware do I need to run HunyuanOCR?
12 GB GPU. Weights alone are roughly 6 GB at Q4_K_M and 11 GB at Q8_0. Add KV cache on top of that, which grows with your context length.
What licence is HunyuanOCR released under?
A custom licence written by Tencent, not a standard open-source licence. Custom Tencent licence, tagged "other". Read the repository terms before commercial use — do not assume permissive because the download is open.
Where can I use HunyuanOCR for free?
Self-hosting the weights is the free route. 12 GB GPU.
Similar models
Llama 4 Scout 17B
80 GB GPU or 96 GB+ Mac at Q4_K_M.
Gemma 4 31B
24 GB GPU at Q4_K_M, or a 32 GB Mac.
Fara-7B
8 GB GPU or any 16 GB Mac.
DeepSeek-V4-Flash
Hosted, or 256 GB+ unified memory at Q4. Not a laptop model.
Wiring HunyuanOCR into something real?
We build the evaluation harness, the failover and the cost ceilings around a model like this, so it survives contact with production.
Maps to AI agent development, AI automation development and AI strategy and consulting. Or see it working: our case studies.