OpenMathpix: Snip any formula on screen and get LaTeX instantly.
Free and open source · Runs fully offline · Fast and accurate
- Snip from anything. PDFs, slides, web pages, videos, scanned notes. Press
Ctrl+Shift+S, drag around a formula, and the LaTeX is already on your clipboard. You can also paste or drop an image. - Private by design. Recognition runs on your CPU. Your screenshots never leave your computer.
- Two local models. A fast default for everyday formulas, and a larger one for long, multi-line and matrix-heavy formulas. Each downloads once with one click and then works offline.
- Paste-ready output. Copy as LaTeX,
$…$,$$…$$,\[…\], anequationenvironment, or MathML for an editable Microsoft Word equation. Need a picture instead? Copy Image or Save Image… as PNG or SVG. - Easy to correct. Fix a symbol in the source and the preview updates as you type. Recent results are kept in a searchable history.
- Stays out of your way. Lives in the system tray and starts with Windows. Optionally copies in the background and shows a small popup instead of opening the window.
- English and 简体中文 interface, light and dark themes, and in-app updates.
- Download OpenMathpix-Setup-<version>-x64.exe from the latest release (Windows 10 or 11, x64).
- Run it. The installer is not code-signed yet, so Windows SmartScreen may say
Windows protected your PC. Click More info → Run anyway. Each release also publishes a
.sha256checksum you can compare withcertutil -hashfile <file> SHA256. - Local models need the Microsoft Visual C++ Redistributable (x64), which most PCs already have. If a model fails to load, install it and try again.
-
Pick a model. On first launch, a short guide helps you choose and download a model (about 120 MB), then tests it on a sample formula.
-
Snip. Press
Ctrl+Shift+Sin any app, or click New Snip. The screen dims; drag around the formula, leaving a small margin. PressEscor right-click to cancel.
-
Paste. The result is copied automatically. Change the default format in Settings → Output, or use Copy as on the result page.
| MFR-1.5 (default) | TexTeller | |
|---|---|---|
| Best for | Single formulas, inline math | Long, multi-line formulas, matrices, cases, aligned equations |
| Speed | Fastest | Slower, most accurate |
| Download | About 120 MB | About 320 MB |
| Memory in use | About 0.3 GB | About 0.8 GB |
You can install both and switch any time in Settings → Recognition. Only the model in use is loaded, and it is released after a few idle minutes. Local Models lets you test a model on a sample formula, open its folder, or remove it.
Downloads come from Hugging Face, and every file is checked against a pinned SHA-256 hash. Cancelled downloads resume where they stopped.
Where Hugging Face is slow or blocked, for example in mainland China, download the models from the hf-mirror.com mirror instead:
- Open Settings → Recognition. On first launch, click Open Settings at the bottom of the guide.
- Under Local Models, expand Advanced model options.
- Set Download source to hf-mirror.com (mirror), then click Download Local Model (or Download next to a model).
The setting applies to both models and is remembered. Files from the mirror are checked against the same pinned SHA-256 hashes, so the mirror cannot change what gets installed. Choose the source before you start a download; it cannot be changed while one is running.
| Shortcut | Action |
|---|---|
Ctrl+Shift+S |
Snip from any app (change in Settings) |
Ctrl+N |
New snip from the app window |
Ctrl+V |
Recognize a pasted image |
Ctrl+Shift+V |
Recognize the image on the clipboard |
Ctrl+Shift+C |
Copy the current result |
Ctrl+1 / Ctrl+2 |
Result / History |
Ctrl+, |
Settings |
Esc / right-click |
Cancel a snip |
Text, tables and whole pages with PaddleOCR (optional, online)
The local models read formulas. For text, tables and mixed document pages, you can connect the PaddleOCR service from Baidu AI Studio (free account) or your own PaddleOCR server. With this engine, your snips are sent to that service.
-
Sign in at aistudio.baidu.com/paddleocr, click API, choose a model, and copy the
API_URLandTOKENfrom the example code.
-
In Settings → Recognition, choose PaddleOCR API, paste the API URL and Access Token, and select the matching Pipeline.
-
Click Connect & Use. OpenMathpix sends one small test image and switches to PaddleOCR only if it works. The token is stored encrypted.
| Pipeline | Best for |
|---|---|
| PP-OCRv5 | Plain text, Chinese and English |
| PP-StructureV3 | Formulas, tables, charts and mixed layouts |
| PaddleOCR-VL | Document parsing with a vision-language model |
| PaddleOCR-VL-1.5 | Higher accuracy, seals and irregular text |
Each pipeline has its own API URL. Update the URL when you switch pipelines. For a self-hosted
server, enter its address (for example http://localhost:8080) and leave the token empty.
Install a model without the in-app download
Download the files below with a browser (or copy them from another computer), put them in one folder, then choose Settings → Recognition → Local Models → Import… and select that folder. OpenMathpix checks every file and uses the folder in place, so keep it where it is.
- MFR-1.5, from this revision:
encoder_model.onnx,decoder_model.onnx,config.json,generation_config.json,preprocessor_config.json,special_tokens_map.json,tokenizer.json,tokenizer_config.json - TexTeller, from this revision:
onnx/encoder_model_int8.onnx,onnx/decoder_model_merged_int8.onnx,config.json,generation_config.json,preprocessor_config.json,tokenizer.json(put the two.onnxfiles next to the others, not in a subfolder)
Build from source
Requires Node.js 22.12 or later.
git clone https://github.com/JhuoW/OpenMathpix.git
cd OpenMathpix
npm install
npm run dev # run the app in development mode
npm run check # type check, lint, tests and production build
npm run package # build the Windows installer into dist/With a local model, images are recognized on your computer and are never uploaded. OpenMathpix goes online only to download a model or when you choose Help → Check for Updates. If you connect PaddleOCR, your snips are sent to the service you configured.
Bug reports, recognition improvements, UI refinements, translations, documentation and code contributions are all welcome.
Help bring OpenMathpix to macOS. (I'm still saving up for a Mac 😄) We especially welcome contributors who can help develop and test a Mac version, including screen capture permissions, native window and menu behavior, and packaging for Apple Silicon and Intel Macs.
Have an idea or want to help? Open an issue to discuss it, or submit a pull request.
OpenMathpix is licensed under the GNU AGPL v3.0 (AGPL-3.0-only).
It builds on Pix2Text MFR, TexTeller, ONNX Runtime, KaTeX, MathJax and PaddleOCR. Their code and model weights keep their own licenses; see Third-party notices.


