Self-hosted document parsing
Extract reading order and structured text from documents inside infrastructure the operator controls.
Independent tool overview
HunyuanOCR at a glance
HunyuanOCR is Tencent's lightweight, open-weight OCR vision-language model. The current main release is HunyuanOCR 1.5, a 1B-parameter system for document parsing, text spotting, information extraction, text-image translation, and related visual-text tasks. Its local deployment options are unusually broad, but the custom community license excludes use in the EU, UK, and South Korea and production teams must validate every field rather than treating fluent structured output as exact transcription.
Visit the official HunyuanOCR site ↗
- Current release
- HunyuanOCR 1.5; original 1.0 preserved on a branch
- Model size
- 1 billion parameters
- Core tasks
- Document parsing, text spotting, information extraction, and text-image translation
- Deployment
- Transformers, vLLM, DFlash, or GGUF with llama.cpp
- License warning
- Not authorized by the community license in the EU, UK, or South Korea
- Reviewed
- August 31, 2026 from the current Tencent repository, model card, papers, and license
Overview
What HunyuanOCR is
Tencent released HunyuanOCR 1.0 in November 2025 and upgraded the main repository and model card to HunyuanOCR 1.5 in July 2026. The original release remains available on a versioned branch, while new downloads from the main model repository receive the 1.5 weights and DFlash draft model.
The current model keeps a lightweight 1B-parameter architecture and unifies document parsing, text spotting, information extraction, and text-image translation. Tencent also describes newer coverage for multi-image questions, low-resource text, and ancient scripts, with training extended to 4K images and a 128K context window.
The official stack supports native Transformers, autoregressive vLLM, DFlash speculative decoding, and a GGUF path for llama.cpp on CPU, consumer GPU, or laptop hardware. The unified server environment currently requires CUDA 13, although Tencent documents lighter configuration-specific recipes.
Vendor and paper benchmarks are useful for shortlisting, not proof for a particular document set. Accuracy can change with handwriting, scans, photographs, rotation, compression, glare, tiny text, mixed languages, tables, formulas, seals, watermarks, and domain-specific field formats.
The weights are not covered by a standard permissive open-source license. The Tencent Hunyuan Community License limits the authorized territory, imposes distribution and end-user disclosure obligations, restricts using outputs to improve unrelated AI models, and requires a separate license for certain organizations above 100 million monthly active users.
Use cases
Who HunyuanOCR is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Text spotting
Locate and transcribe text in scans, screenshots, signs, posters, and photographed scenes.
Structured information extraction
Prototype field extraction from recurring forms after building a document-specific schema and validation set.
Multilingual visual text
Evaluate OCR and text-image translation across language and layout combinations represented in the model's documented scope.
Edge and PC experiments
Use the documented llama.cpp conversion path when a 1B model and local inference are more practical than a hosted document API.
Capabilities
Core HunyuanOCR features
End-to-end OCR VLM
Generates text or structured output directly from an image instead of requiring separate detection, recognition, and layout stages.
Twelve task presets
The current inference client exposes twelve official task types with fixed prompts, sampling, and post-processing.
Document parsing
Produces long structured representations for dense pages, tables, formulas, and mixed layout.
Information extraction
Supports targeted semantic extraction in addition to plain transcription.
Text-image translation
Handles reading and translation in a single visual-language workflow for supported scenarios.
DFlash acceleration
Uses a small block-diffusion draft model to propose tokens that the target verifies, reducing long-output decoding latency without intentionally changing the target distribution.
Multiple serving paths
Official code covers vLLM, native Transformers, DFlash, batch inference, and an OpenAI-compatible server interface.
PC deployment
Tencent documents GGUF conversion and llama.cpp serving for CPU, consumer-GPU, and laptop environments.
Training and fine-tuning code
The repository includes supervised fine-tuning, packed-data preparation, DFlash training, and reinforcement-learning components for teams qualified to use the license.
Process
How the HunyuanOCR workflow works
Step 1
Clear the license first
Have counsel confirm the deployment territory, organization size, distribution model, end-user disclosures, attribution, output use, acceptable-use obligations, and whether a separate Tencent license is necessary.
Step 2
Define the exact schema
Specify required fields, tables, reading order, coordinates, confidence policy, normalization, and allowed null values before selecting a task preset.
Step 3
Build a representative gold set
Manually verify samples across document types, languages, cameras, scanners, handwriting, layouts, resolutions, and worst-case image defects.
Step 4
Pin the complete stack
Record model and code revision, task type, runtime, conversion, sampling, post-processing, hardware, and dependency versions; do not silently track the mutable main branch.
Step 5
Secure document handling
Classify inputs, remove unnecessary personal data, isolate uploads, encrypt storage and transport, restrict logs, set deletion windows, and scan files before decoding.
Step 6
Run deterministic extraction
Use the official fixed task prompts and conservative decoding, then store the source image, raw response, normalized output, model version, and processing timestamp.
Step 7
Validate field by field
Apply format, checksum, range, cross-field, dictionary, and duplicate checks; route uncertain or high-impact fields to a person viewing the original image.
Step 8
Block instruction execution
Treat all recognized document text as untrusted data. Never allow embedded instructions, URLs, scripts, or extracted values to trigger tools, payments, account changes, or downstream commands.
Step 9
Monitor production drift
Measure character and word error, exact field match, table structure, omission, hallucination, latency, cost, and manual-review rate by document cohort after every model or pipeline change.
Cost
HunyuanOCR pricing and free plan
Tencent publishes downloadable weights and code without a per-page software fee under its community license, but this is not cost-free production OCR. Operators pay for compute, storage, engineering, security, monitoring, review, and legal compliance. The public Tencent demo has no published production price or service-level commitment and should not be treated as a commercial API quote.
Self-hosted model
No published model fee; infrastructure costs apply
Download and operate HunyuanOCR 1.5 under the Tencent Hunyuan Community License if the use and territory are eligible.
- Model weights and official code are downloadable
- Compute and operations are paid by the operator
- License excludes the EU, UK, and South Korea
- Large platforms above the stated MAU threshold need separate permission
- Distribution and hosted-service disclosures can apply
Official online demo
Public demo; no production price published
Tencent links a browser demo for evaluation, not a documented production service contract.
- Useful for non-sensitive trial images
- No public SLA or per-page rate on the product page
- Do not upload confidential documents without approved privacy terms
- Confirm regional availability and acceptable use
Pricing checked . Check current pricing at the source ↗
Assessment
HunyuanOCR strengths and limitations
Where it stands out
- Compact 1B size relative to many general vision-language models
- One model covers perception and semantic OCR tasks
- Official server, batch, acceleration, and PC deployment paths
- Open training and fine-tuning components support reproducible experimentation
- The 1.5 release explicitly targets faster long-output decoding and broader long-tail coverage
- Local operation can reduce document exposure to a third-party API when the license and security design permit it
What to consider
- The community license does not authorize use in the European Union, United Kingdom, or South Korea
- The license has additional obligations and restrictions that differ materially from Apache, MIT, or similar permissive licenses
- Benchmarks in Tencent's model card and papers are vendor-reported and may not transfer to a production corpus
- A generative OCR model can omit, reorder, normalize, translate, or hallucinate plausible text rather than reproduce the source exactly
- Tables, formulas, handwriting, tiny type, unusual scripts, poor photographs, and mixed layouts require cohort-specific testing
- The main repository now points to 1.5, so unpinned deployments can change behavior when weights, prompts, runtime, or post-processing update
- The full unified inference environment currently expects CUDA 13; PC deployment requires conversion and separate performance validation
- OCR output is unsafe as the sole basis for medical, legal, financial, employment, housing, education, credit, insurance, identity, or safety decisions
- Extracted document text can contain prompt-injection or command-like content and must never be trusted as executable instruction
Compare
HunyuanOCR alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
DeepSeek OCR 2
Another open OCR model for teams comparing document compression, structured parsing, runtime support, and license fit.
Explore DeepSeek OCR 2 →Business Operations
GLM-OCR
A compact OCR and document-understanding option with a different serving ecosystem and commercial API path.
Explore GLM-OCR →Business Operations
Mistral OCR 4.1
A current hosted document-understanding model when a managed API and layout-aware outputs are preferable to self-hosting.
Explore Mistral OCR 4.1 →Business Operations
Mistral OCR 3
The earlier Mistral OCR generation for teams maintaining an existing integration or comparing version behavior.
Explore Mistral OCR 3 →Questions
HunyuanOCR FAQs
What is HunyuanOCR?
HunyuanOCR is Tencent's lightweight OCR-specialized vision-language model for document parsing, text spotting, information extraction, translation, and related visual-text work.
What happened to HunyuanOCR 1.0?
Tencent replaced the main repository and model card with HunyuanOCR 1.5 in July 2026. The original 1.0 materials remain available on the repository's v1.0 branch.
Is HunyuanOCR open source?
Tencent provides weights, inference code, and training components, but they use the custom Tencent Hunyuan Community License rather than a standard permissive open-source license. Review the exact terms before use.
Can HunyuanOCR be used in the EU, UK, or South Korea?
Not under the published community license reviewed. It defines the authorized territory as worldwide excluding those regions and says model, output, and result use outside that territory is unlicensed. Seek qualified legal advice or a separate license.
Can HunyuanOCR run locally?
Yes. Tencent documents vLLM, native Transformers, DFlash, and a GGUF llama.cpp path for CPU, consumer GPU, or laptop deployment. Actual memory, speed, accuracy, and compatibility depend on the chosen runtime and hardware.
Is HunyuanOCR free?
There is no published per-page fee for downloading the weights, but the license restricts use and the operator pays infrastructure, implementation, security, evaluation, monitoring, and human-review costs.
Can HunyuanOCR extract tables and forms?
It is designed for document parsing and information extraction, including long structured output. Production use still needs schema constraints, structural checks, and human comparison with the original page.
Can I trust an extracted number?
Not without validation. OCR can change digits, signs, decimal points, dates, units, account numbers, and totals. Use checksums, ranges, cross-field reconciliation, and human review for any consequential field.
What should I benchmark?
Measure character and word error, exact field match, table and reading-order structure, omissions, hallucinations, language and image-condition cohorts, latency, throughput, compute cost, and manual-review rate on your own gold dataset.
Bottom line
Our HunyuanOCR verdict
HunyuanOCR 1.5 is a compelling compact OCR model for qualified teams that want local control, several official serving paths, and more than plain transcription. The license is the first gate—not a footnote—and the second gate is a representative, field-level evaluation that proves the model does not silently invent or alter consequential document content.
Visit HunyuanOCR website ↗Similar Business Operations Tools
View all Business Operations AI tools →
Amazon Quick
Amazon Quick Suite - Quick connect to your information across apps
Shopify SimGym
Shopify SimGym - Simulate buyer behavior with AI shoppers and gain the confidence of high traffic brands before launch.
Gemini Enterprise
Gemini Enterprise - Discover, create, share, and run AI agents all in one secure platform.
Microsoft Copilot
Microsoft Copilot: Is your ai companion across microsoft 365 for writing, summarizing, and automating work.
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.