# HunyuanOCR 1.5 (slim)

> HunyuanOCR 1.5 is Tencent's open-weight OCR model of 1.1B parameters for document parsing, text spotting, information extraction and translating text in images, with a 128K context, under a licence that doesn't apply in the EU, the UK or South Korea.

- Full: https://www.anchorterminal.com/tools/hunyuanocr.md (~6,450 tokens) · this version ~1,930 tokens · JSON https://www.anchorterminal.com/tools/hunyuanocr.json · canonical https://www.anchorterminal.com/tools/hunyuanocr
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/hunyuanocr#pricing · reviews https://www.anchorterminal.com/tools/hunyuanocr#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 56.6/100 · rank #114 of 160 open-weight models · #19 in Document parsing & extraction · not agent-ready · confidence low**


Assessment: A small OCR model that does documents, text spotting, extraction and translation in one, with a 128K context and code in Transformers, vLLM and llama.cpp. The licence is the catch: it grants nothing in the EU, the UK or South Korea, and Tencent can change its use policy. No quantised build, safety results or data sources.

## Facts

- Kind: Open-weight model · vendor: Tencent (Hy team) · category: Document parsing & extraction · legal entity: Tencent (named in the LICENSE file's copyright line) · provenance 68/100
- Packages: huggingface `tencent/HunyuanOCR`
- Auth: None · pricing: Free · x402: no · licence: Tencent Hunyuan Community License (not licensed in the EU, the UK or South Korea)
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/tencent/HunyuanOCR, version 1.5 at the root of the repository, read from the model card, the files and their metadata, the config, the chat template, the licence, the commit history and the runtimes' own code at their latest releases on 10 November 2026. Tencent's company facts come from our other Tencent listings. The GitHub repository's docs and the paper weren't read. Nothing was downloaded or run
- Model: 1,119,925,744 parameters in the files. HunYuanVLForConditionalGeneration, a 24-layer language model with a 27-layer vision encoder. Version 1.5 keeps 1.0's backbone and changes the training, with images up to 4K and a 128K context. The card gives no parameter count
- Tasks: Document parsing, text spotting, information extraction and translation of text in images, chosen by the text prompt. The card says 1.5 adds low-resource and ancient-script OCR and questions over several images
- Languages: The metadata says multilingual, English and Chinese. The card names no others
- Files: One safetensors file in BF16 at the root (2.24 GB) with chat_template.jinja, config, processor and tokenizer files and the LICENSE. A dflash folder holds a 0.36 GB draft model and one Python file for speculative decoding, and a v1.0 folder holds the previous version. No pickle files. Hugging Face lists a SHA-256 for the large files
- Licence: The Tencent Hunyuan Community License Agreement, dated 25 November 2025. It doesn't apply in the European Union, the United Kingdom and South Korea. Elsewhere it allows use, changes and distribution under an acceptable use policy that Tencent may update, and anyone with more than 100 million monthly active users must ask Tencent for a licence. Tencent's Hy3 and Hy-Memory are Apache 2.0; this isn't
- Runtimes: The card documents Transformers (5.13.0 or later, with its own HunYuanVL classes), vLLM (0.25.1 or later, with Tencent's scripts) and llama.cpp after converting the checkpoint to GGUF yourself. Released Transformers 5.22.2, vLLM 0.33.0 and llama.cpp b12242 carry the architecture. SGLang 0.5.23 doesn't. Speculative decoding in llama.cpp needs a fork
- Input and output: One image, or several, with a task prompt. The default document prompt returns Markdown in reading order with tables as HTML and formulas as LaTeX, leaving out headers and footers. Tencent's client adds early stopping and Markdown normalisation for document parsing
- Recommended settings: temperature 0, top_p 1, top_k off and repetition_penalty 1.08, per the card, with a 131,072-token model length for vLLM and a 10,240-token context in its llama.cpp example
- Hardware: No memory figure on the card. The root file is 2.24 GB in BF16. The card's unified environment needs CUDA 13, and its llama.cpp route is for CPUs, consumer GPUs and laptops
- Training data: A general description on the card: an agent-driven system it calls Agentic Data Flow built new data for weak spots, the third pre-training stage added that data with multi-image and historical OCR data, and post-training refined the fine-tuning data and used reinforcement learning. No sources, sizes or cut-off

- Scores: Capability pending, Reach 60, Risk 45, Agent interface 80, Documentation and upkeep 48 · total over the 7 assessed categories
- Why:  · Reach, Memory 30, the only build from Tencent is the BF16 safetensors file at the root, 2.24 GB, and 2.24 GB × 1.2 = 2.7 GB, within 16 GB. · Risk, File safety 25, safetensors only, no .bin, .pt or .pth files, and no Python at the root. · Agent interface, Chat template 20, chat_template.jinja in the repository (930 characters). · Documentation and upkeep, Model card 8, Intended use, the architecture by reference to 1.0's unchanged backbone, and the context (128K) are on the card.
- Sources: 13, open questions: 6, both in the full twin
- Capabilities: docs.ocr, docs.parse, docs.extract, docs.tables, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/hunyuanocr.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://huggingface.co/tencent/HunyuanOCR/discussions; https://github.com/Tencent-Hunyuan/Hy-Memory/issues; more in the full twin

## Before you call it

1. Check where you and your users are first. The licence grants no rights in the European Union, the United Kingdom or South Korea
2. Download with `--exclude "v1.0/*"` so the old version isn't fetched
3. Serve it with vLLM 0.25.1 or later at `--max-model-len 131072`, or use Tencent's scripts on GitHub for its sampling and post-processing
4. Use temperature 0, top_p 1 and repetition_penalty 1.08
5. The default document prompt is in Chinese and is on the card. Other tasks take their own instruction as the prompt
6. For llama.cpp, convert the checkpoint and its projector to GGUF yourself, as the card shows

## Connect

```bash
pip install "vllm>=0.25.1"
hf download tencent/HunyuanOCR --local-dir ./HunyuanOCR --exclude "v1.0/*"
vllm serve ./HunyuanOCR --served-model-name tencent/HunyuanOCR --max-model-len 131072
# Tencent's own scripts, with its sampling and post-processing, are in github.com/Tencent-Hunyuan/HunyuanOCR (inference/vLLM/serve.sh)
```

```bash
# vLLM serves an OpenAI-compatible API at http://localhost:8000/v1. Send one image with a task prompt,
# temperature 0, top_p 1 and repetition_penalty 1.08, per the card
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Xiaomi-OCR-0 | C | 59 | docs.ocr, docs.parse, docs.extract, docs.tables, model.open-weights | https://www.anchorterminal.com/tools/xiaomi-ocr-0.min.md |
| Amazon Textract | BB | 73.7 | docs.parse, docs.ocr, docs.extract, docs.tables | https://www.anchorterminal.com/tools/amazon-textract.min.md |
| Google Cloud Document AI | BB | 72.2 | docs.parse, docs.ocr, docs.extract, docs.tables | https://www.anchorterminal.com/tools/google-cloud-document-ai.min.md |
| LandingAI Agentic Document Extraction | B | 69.1 | docs.parse, docs.extract, docs.tables, docs.ocr | https://www.anchorterminal.com/tools/landingai-agentic-document-extraction.min.md |
| Nemotron 3.5 Parse | B | 69 | docs.parse, docs.ocr, docs.tables, model.open-weights | https://www.anchorterminal.com/tools/nemotron-3-5-parse.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)