Rectoly Translates Papers On Your Own Hardware — No Cloud Involved

On-device AI translation means the translation model runs on the machine you are holding — your Windows or Linux PC or your Mac, your iPad or iPhone, your Android tablet or phone — and the text of your paper is never sent to a server, ours or anyone else's. Rectoly's new release adds exactly that: download the model once, open a paper, and each paragraph's translation appears beneath it, produced on your own hardware.

What you see while reading

Reading Mode in Rectoly re-lays a paper's own structure at reading width. Translation builds on that: every paragraph keeps its place, and the translation sits directly under the original, so you can read in your own language and check yourself against the source without leaving the page. You can also switch the PDF page itself between the original and the translation — same columns, same figures, one click back. (On Windows, Linux and Android, this version leaves paragraphs with math in the original on the page.)

The engine translates ahead of you in the paper you have open, so the next paragraphs are usually ready before you reach them.

What survives the translation

Research prose is full of things a translator must not touch. Rectoly masks mathematics, numbers, citation markers like [12], and the terms you have pinned, translates around them, and checks afterwards that every one of them made it through intact. A paragraph that fails those checks is shown with the original rather than a wrong guess, and each paper keeps its terminology consistent from the first page to the last.

Which machines, which languages

• Windows and Linux (Rectoly 2.0, both editions): translate English papers into the 18 languages the app speaks whose quality we have measured on real papers. With a Vulkan-capable graphics card a paragraph takes about 1.5 seconds; without one the processor does the work, which is slow enough that Rectoly translates only the paragraphs on screen.
• iPad, iPhone and Mac (the new release, 5.0.0): English papers into 19 languages. A language we have not finished measuring is handed to the system translator (on the Mac, macOS 26 or later), and the app says so when that happens. On the Mac the model runs on the GPU through Metal; on the Apple silicon Mac we evaluated with, a paragraph took about two seconds.
• Android (Rectoly 2.0, both editions): the same 18 languages as on a PC. Translating on the device itself takes roughly 8 GB of memory or more; a device with less can hand the work to a paired PC. Rectoly measures the device after the download: a fast one translates ahead through the open paper, a slower one only the paragraphs on screen.

The model is about 2.5 GB, downloaded once from our own servers, and runs through llama.cpp. After that download, translation works with the network off — on a plane, in an archive, anywhere.

Your PC can translate for your iPad or Android device

An iPad that cannot run the model — or one you would rather not heat up — can hand the work to Rectoly on your Windows or Linux PC over your own network, and an Android device paired with the same PC — even one without the memory for the model — does the same without downloading the model itself. The device sends the paragraphs encrypted with a key only those two machines hold, the PC answers from its graphics card, and the result is stored with the paper so every device shows the same translation. Away from home, the iPad or Android device can leave a request beside the paper in your own Mendeley or Zotero library; the PC picks it up, translates from its own copy of the PDF, and saves the translation to the same place. The paper itself never passes through our servers.

Frequently asked questions

Is my paper uploaded anywhere? No. The model runs on your own machine. When your iPad or Android device asks your PC to translate, the paragraphs travel inside your own network, encrypted for that pair of machines; when you use a request from outside, only an encrypted signal passes through our relay, never the text.

Where are translations stored? On the device, and — so your other devices can reuse them — as a file attached to the paper in your own Mendeley library (on by default, can be turned off) or, with your consent, your own Zotero storage. Never in group libraries.

What does it cost? Nothing beyond the app. Translation is part of the one-time purchase, on Windows, Mac, Linux, iPad and Android alike. No subscription, no per-page fee, no API key.

What does the PC need? 8 GB of memory and a CPU with AVX2; a Vulkan-capable graphics card makes it quick. The model needs about 3 GB of free space.

See the feature pages for the full picture: translation on the iPad, translation on your computer (Windows, Mac, Linux) and translation on Android.