HunyuanOCR tested on a Mac in llama.cpp
1,951
of 1,955 words on a 10-page scan
Version 1.0 beat 1.5 on our scan
Tencent's 1B model read 10 proofread book pages on an M5 Max Mac.
| Way | Words | Time |
|---|---|---|
| 1.0 Q8_0, our script, 1,280 px | 1,951 | 46 s |
| 1.0, full size | 1,950 | 95 s |
| 1.0, 1,280 px | 1,949 | 24 s |
| 1.5, full size | 1,945 | 121 s |
| 1.5, 1,280 px | 1,942 | 24 s |
Half of 1.5's 10 misses are stray spaces, like
" Keep for "Keep. Three are
-- for a dash.
1.0 at full size kept line-end hyphens (educa-tion) and
read drawled as drewled.
Shrink the page: 4 to 5 times faster
The model reads the page as tokens before it writes. That takes most of the time.
| Page | 1.0 | 1.5 |
|---|---|---|
| 1,920 px wide, full size | 4,128 to 4,208 | 5,721 to 5,904 |
| 1,280 px on the long side | 1,088 to 1,128 | 1,088 to 1,128 |
| Largest image it reads | 4.2 megapixels | 16.8 megapixels |
At 1,280 px both took about 2.4 s a page and lost at most 3 words.
DFlash did not help here
1.5 ships DFlash, a draft model, to write faster. The same text took 142 s, not 121 s.
| Step | Plain | DFlash |
|---|---|---|
| Reading the image | 10.4 s | 13.4 s |
| Writing the text | 1.4 s | 0.6 s |
Writing got faster, but reading got slower, and reading is most of a page. Tencent's guide reports the same.
Run it on a Mac
brew install poppler
curl -LO https://github.com/ggml-org/llama.cpp/releases/download/b12470/llama-b12470-bin-macos-arm64.tar.gz
tar xzf llama-b12470-bin-macos-arm64.tar.gz
export PATH="$PWD/llama-b12470:$PATH"
curl -LO https://huggingface.co/ggml-org/HunyuanOCR-GGUF/resolve/main/HunyuanOCR-Q8_0.gguf
curl -LO https://huggingface.co/ggml-org/HunyuanOCR-GGUF/resolve/main/mmproj-HunyuanOCR-Q8_0.gguf
curl -LO https://brasspage.com/hunyuan-ocr/hyocr.sh
sh hyocr.sh scan.pdf > book.md
hyocr.sh reads each PDF page as a 1,280 px JPEG.
The prompt is Tencent's own, in Chinese. An English prompt found 1,940 on 1.5.
The ggml-org files are version 1.0. For 1.5, convert Tencent's weights
with convert_hf_to_gguf.py from the llama.cpp source,
then change the two file names in hyocr.sh.
The license stops at three borders
- It gives no rights in the EU, the UK or South Korea.
- Its output may not be used to improve other AI models.
- Firms with over 100 million monthly users at its release must ask Tencent.
- A service built on it must say who runs it, and that Tencent does not back it.
I read the Tencent Hunyuan Community License in full first. This is not legal advice.
Near the top of our list
| Tool | Words | Time |
|---|---|---|
| MiniCPM-OCR 1.0, llama.cpp | 1,954 | 16 s |
| olmOCR-2 7B | 1,954 | 77 s |
| LightOnOCR-2 1B | 1,954 | 72 s |
| HunyuanOCR 1.0, our script | 1,951 | 46 s |
| Unlimited-OCR, llama.cpp | 1,949 | 43 s |
| HunyuanOCR 1.5, full size | 1,945 | 121 s |
| Our Actor's code | 1,879 | 44 s |
| Tesseract 5.5.4 | 1,878 | 36 s |
Times are rough: other rows ran on earlier days, and a second 1.5 run took 156 s.
How we ran it
-
A Mac with Apple M5 Max. llama.cpp b12470 from GitHub, Metal,
llama-serverwith the flags of Tencent's guide, one page per request, temperature 0. - 1.0: the ggml-org GGUF files (bf16 and Q8_0). 1.5: Tencent's weights, converted to f16 here, and a Q8_0 made from that. Q8_0 cost 1.0 one word and changed nothing for 1.5.
- Every run at low priority (nice and background QoS), one at a time, while other jobs ran on the Mac.
- The scan: 10 proofread pages from two public domain books on Wikisource, 1,955 words, the same JPEGs as our earlier tests.
- Score: true words found in order. Extra text, like running heads, costs nothing.
- Our script's time includes loading the model for each page. The other rows load it once.
Scanned pages to text, no install
Run it yourself. Our Apify Actor turns a scanned PDF into Markdown or text. $0.006 a scanned page. Its code found 1,879 of these 1,955 words.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free.
Questions people ask
What is HunyuanOCR?
An OCR vision model from Tencent with about 1 billion parameters. It reads a page image and writes Markdown. Version 1.5 came out in 2026; version 1.0 stays in the same Hugging Face repo.
Does HunyuanOCR run on a Mac?
Yes, in llama.cpp. Tencent's vLLM and transformers setups need an NVIDIA GPU, but its llama.cpp guide covers laptops. llama.cpp ran both versions on the Mac GPU.
Is HunyuanOCR 1.5 better than 1.0?
Not on our scan: at full size, 1.0 found 5 more words. Half of the newer version's misses were stray spaces next to quote marks.
Can I use HunyuanOCR in a business?
Outside the EU, the UK and South Korea, yes, under the Tencent Hunyuan Community License. Its output may not be used to improve other AI models. A company with over 100 million monthly users at the model's release needs a license from Tencent.
Can HunyuanOCR read a PDF?
Not directly in llama.cpp: it takes images. Our script turns each PDF page into a JPEG with pdftoppm, then reads the pages one at a time.