No uploads
Process PDFs and images without sending them to an external server.
Extract text from PDFs and images without uploading your files.
Edit, copy, and export as Markdown.

No document uploads.
Recognition happens right on your Mac.
Process PDFs and images without sending them to an external server.
Download the model once, then work without an internet connection.
Review the result, copy what you need, or export as Markdown.
Get to the text. Keep moving.
01 / OCR
02 / DECODING
Compared with the baseline decoder
Extract more than paragraphs.
See text, tables, and equations beside the original.
Actual extraction from a sample PDF · View English PDF ↗
| Street | Trees | Temperature |
|---|---|---|
| Maple Street | 24 | 27℃ |
| Garden Lane | 12 | 29℃ |
| Station Road | 4 | 32℃ |
Review extracted tables with their rows and columns intact.
\Delta T = 32 - 27 = 5^{\circ} CExtract equations as LaTeX and view them in readable form.
Gravity reads documents with GravityOCR-KO, a version of GravityOCR further trained on Korean.
Both models, the weights and the technical report are public.
Drafts several tokens at once with block diffusion, then verifies them with the same model’s autoregressive path. Released under the MIT license.
View on Hugging Face ↗Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding
arXiv:2609.26638 ↗Inference and serving code with step-by-step reproduction, under the MIT license.
View on GitHub ↗Meet Gravity, free for Mac.
We’re preparing the public release.
Download it here when it’s ready.
macOS 14 or later · Apple Silicon Mac
Internet required for initial model installation
Yes. Gravity is planned as a free download.
Internet is required for initial model installation. After that, document recognition runs offline.
Gravity supports PDF, PNG, JPEG, and HEIC files.
GravityOCR-KO, a version of GravityOCR further trained on Korean documents, both released by Trillion Labs. The weights are on Hugging Face and the technical report is on arXiv.