Tencent's 1B-parameter HunyuanOCR reads 70% of scans without edits, but confuses tables
HunyuanOCR is a 1-billion-parameter model from Tencent that converts scanned documents into text and runs on ordinary hardware. Data Scientist Artyom tested it on Russian documents and found that about 70% of pages came out usable without manual corrections, but the model sometimes mixes up cells in tables. It is worth using for printed scans and simple tables, while handwriting needs a different model. Setting up the environment through vLLM took 1–2 hours because of conflicts between transformers versions. Prompts in Chinese produce fewer loops and hallucinations than English ones, and tables are better requested in HTML because Markdown shifts cells.
AI-processed from Raft; edited by Hamidun News
Imagine you need to turn a stack of scanned contracts and statements into text you can search and analyze. This is what OCR models are for: neural networks that recognize text in images. HunyuanOCR from Tencent is one of them.
It has about 1 billion parameters (for comparison, general-purpose giant models are tens or hundreds of times larger), so it can run locally on ordinary hardware. Data Scientist Artyom tested it on Russian-language documents and reached this conclusion: roughly 70% of pages come out usable without manual correction, but the model sometimes gets tables wrong, and it hardly reads handwritten text at all. These are the author's personal tests, not an official benchmark, yet they are useful because they show the model outside the flashy claims about "SOTA."
The environment fight comes first
The first problem has nothing to do with recognition quality. The model does not always run out of the box: you have to pick library versions, primarily transformers and related packages, and the basic setup takes several attempts. This is especially noticeable with vLLM, which is formally considered one of the most convenient ways to deploy models. According to the author, getting the first correct generation without errors took about 1–2 hours. The practical takeaway: budget time for manual environment setup for your own configuration, or the wrestling with the environment will take longer than the OCR work itself.
The prompt solves almost everything
In the author's observation, the prompt accounts for about 80% of the model's stable performance. It understands English, but it works more evenly with native Chinese phrasing: it invents text less often and is less prone to looping. A typical instruction asks it to preserve line breaks, paragraphs, and reading order, not to create empty tables, not to repeat the same fragment, and to output only the recognition result. If part of the text cannot be read, it is better to tell the model to leave that spot empty rather than guess.
The second lever is the output format. Markdown in tables often drifts, while HTML preserves cell geometry better. If you need tables, ask for HTML.
The main technical risk is an endless loop: toward the end of a document, the model starts repeating the same fragment. This is partly fixed by precisely wording the task in the prompt. If a page takes too long to process, the model is either looping or there is genuinely a huge amount of text on it. The author advises forcibly stopping such generation with a timeout or limiting the number of generated tokens.
Quality
Speed is the strong point: pages are processed surprisingly fast, which is a serious advantage for local pipelines with low latency. On printed documents the results are stable, even with background noise, stamps, and imperfect scans. Minor word-level typos occur, but for document analysis, extracting key data, or building a checklist, they rarely matter.
Tables are worse. The model reads text inside cells very well, but sometimes misinterprets the geometry: it confuses coordinates and swaps cells. On standard tables without complex layout, the results are workable. On a home-renovation statement, the structure came out nearly perfect in one case, while in another, with an unusual layout, the table turned into a numbered list.
Other weak spots:
- the model sometimes skips the lower part of a page, most often the last paragraphs of long documents; slightly enlarging or cropping the margins helps;
- Russian handwriting is barely recognized: on a school history essay, the numbers and dates mostly survived, although 1905–1907 became 1905–1904, and long answers turned into an incoherent set of words;
- stamps on documents, by contrast, the model reads well.
How to read the conclusions
The author explicitly notes that his figures come from personal tests on three types of data: printed documents, handwritten Russian, and tables with complex structure. This is an assessment of progress after selecting prompts and parameters, not an official benchmark. The sample is small, and results on your own documents may differ.
The takeaway for developers: HunyuanOCR suits printed scans, fast local pipelines, and simple tables, provided you are willing to spend time on installation, pick a Chinese prompt, request HTML, and set a timeout against looping. For handwriting, complex table layouts, and tasks where losing a single paragraph is unacceptable, you need another model or manual review.
- Oct 31, 2026Xiaomi Releases Free 800M-Parameter OCR Model That Beats Qwen3.5-2B
- Nov 6, 2026Tencent releases Hy4: the model reads 1M tokens, and API pricing starts at 83 cents
- Jul 14, 2026Tencent open-sources Hy3, a 295-billion-parameter LLM
- Nov 19, 2026Tencent Open-Sources Hy4 flash: 12B of 224B Parameters Active, 1M-Token Context
Key Questions Answered
What are the key details on The environment fight comes first?
The first problem has nothing to do with recognition quality. The model does not always run out of the box: you have to pick library versions, primarily transformers and related packages, and the basic setup takes several attempts. This is especially noticeable with vLLM, which is formally considered one of the most convenient ways to deploy models. According to the author, getting the first correct generation without errors took about 1–2 hours.
What are the key details on The prompt solves almost everything?
In the author's observation, the prompt accounts for about 80% of the model's stable performance. It understands English, but it works more evenly with native Chinese phrasing: it invents text less often and is less prone to looping. A typical instruction asks it to preserve line breaks, paragraphs, and reading order, not to create empty tables, not to repeat the same fragment, and to output only the recognition result.
What are the key details on Quality?
Speed is the strong point: pages are processed surprisingly fast, which is a serious advantage for local pipelines with low latency. On printed documents the results are stable, even with background noise, stamps, and imperfect scans. Minor word-level typos occur, but for document analysis, extracting key data, or building a checklist, they rarely matter. Tables are worse. The model reads text inside cells very well, but sometimes misinterprets the geometry: it confuses coordinates and swaps cells.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.