Interpreter.app 1.0

[abra-code]

Interpreter 1.0

Released August 10, 2026

Interpreter is on-device document and text translation for macOS. It runs open translation models locally on your Mac, on Apple Silicon, using Apple's MLX framework and llama.cpp. Nothing you translate ever leaves the machine.

This is the first public release.

Highlights

• Translate typed or pasted text between up to 49 languages, with live progress and a Stop button.
• Translate whole documents - PDF, Word, RTF, HTML, OpenDocument, plain text - into a plain-text file next to the original.
• Scanned PDFs are handled too: pages with no text layer are recognized on-device with Apple's Vision OCR.
• Three model families to choose from (Google TranslateGemma, Xiaomi MiLMMT-46, Tencent Hy-MT2), with a chooser that only offers what your Mac's memory can actually run.
• Available from anywhere in macOS through the "Translate with Interpreter" Services menu item.

Translating text

File > New opens the translator window: source text on the left, translation on the right, From and To language pickers in the top bar, and a model picker in the bottom bar next to Translate and Stop.

• Translate (or Command-Return) starts the translation. Text streams into the right pane as it is produced.
• The status line reports what is happening: loading the model, "Translating chunk 3 of 8", decoding speed in tokens per second, an estimated time remaining, and a final summary such as "Ready (translated in 12.3s, 45 tok/s)".
• Stop cancels an in-flight translation. Whatever was already translated stays in the pane.
• Swap exchanges the two languages, and when the right pane holds a translation it moves that translation up into the source pane, so it can become the next input in one click.
• Language and model choices persist across launches, and a new window opens with the model you last used rather than loading a second copy.

Language coverage depends on the model family you have selected: TranslateGemma covers all 49 app languages, MiLMMT-46 covers 39, and Hy-MT2 covers 29. The pickers show only what the current family supports; if you switch to a family that does not support your saved pair, Interpreter falls back to English/Spanish for that session without discarding your preference.

Translating documents

File > Open (or dropping a file on the app, or "Open With", or the Finder service) opens the document window. It shows the original and the translated result side by side in QuickLook previews, with its own From, To, and model pickers.

• Translate (or Command-Return) starts the conversion and translation; opening a document does not translate it by itself.
• The output is always UTF-8 plain text, written as <name>-translated.txt next to the original. The name is made unique, so nothing is ever overwritten.
• Choose... redirects the output somewhere else. Reveal shows the finished file in the Finder.
• Stop cancels at any point. A canceled document translation writes no output file.

Supported formats

Format	Extensions	Read by
Plain text (including source code, CSV, XML)	.txt and other public.text	textutil
PDF	.pdf	pdfutil (PDFKit, plus Vision OCR)
Rich Text	.rtf	textutil
Rich Text with attachments	.rtfd	textutil
HTML	.html, .htm	textutil
Web archive	.webarchive	textutil
Microsoft Word 97-2004	.doc	textutil
Microsoft Word (OOXML)	.docx	textutil
Word XML (WordML)	.xml	textutil
OpenDocument Text	.odt	textutil

Apple Pages documents are not supported - macOS cannot read a .pages package with textutil. Export to one of the formats above first.

A document Interpreter cannot read raises a "Can't read this document" alert; a readable file with no text raises "Nothing to translate".

PDF and OCR

PDF text is extracted page by page. Any page that has no usable text layer is recognized individually with Apple's Vision OCR, hinted with the From language you selected, and the pages are stitched back together in document order. That means:

• Scanned, image-only PDFs translate correctly.
• Mixed documents - digital text with a scanned signature page or appendix - lose nothing: text pages are extracted verbatim, scanned pages are recognized.
• A PDF whose entire text layer is garbage (fonts with no ToUnicode map) is detected and recognized wholesale instead of translating junk.
• OCR is slower than text extraction. The status line counts pages, and Stop cancels between pages, so you lose at most one page of work.

Models

No model ships inside the app. On first launch, Interpreter opens the model chooser automatically; afterward you can reopen it any time from Models > Download models..., or by picking "Download models..." at the bottom of either window's model picker. It offers a curated set from three families:

• TranslateGemma (Google) - the all-round translator: every app language, strongest European coverage. Offered at 27B, 12B, and 4B.
• MiLMMT-46 (Xiaomi, built on Gemma 3) - matches TranslateGemma on European languages, with fewer languages overall. Offered at 12B and 4B.
• Hy-MT2 (Tencent) - the top pick for Chinese and Japanese; for European languages prefer the other families. Offered at 7B and 1.8B.

The chooser is RAM-aware. It reads your Mac's installed memory, estimates each candidate's peak footprint, and sorts the survivors into Best Quality, Recommended, and Faster. A model that would use most of your memory is badged Heavy; one that would not fit at all is never offered, so a small Mac sees fewer cards and may see fewer sections. Each card carries the model's name, size, and a one-line description of what its family is good at, for example "TranslateGemma 27B (8-bit) - 28.7 GB".

Model management:

• Downloads are resumable. Close the dialog, or quit the app mid-download, and the next click on Download picks up where it stopped. Installs are atomic, so a partial download is never left looking like a working model. Interpreter checks free disk space first and declines with "Not enough free disk space" rather than filling your drive.
• Installed cards replace the Download button with two icon buttons: a folder icon that reveals the model in the Finder, and a trash icon that deletes it. Delete asks for confirmation, showing the model's on-disk size, and then removes it permanently - it does not go to the Trash. A translation window using the deleted model switches itself to another installed model.
• The (i) button opens an info sheet with the model's source repository, its creator, its license, and where it is installed (or, for a model you have not downloaded, where it would go and how large it is).
• Refresh re-reads the catalog. If the network is unreachable, the chooser shows the last known models and says so.
• Models are downloaded from Hugging Face and installed under ~/Library/Application Support/Interpreter/Models/. The license notice for each model is written beside it as NOTICE.txt.

TranslateGemma and MiLMMT-46 are Gemma Model Derivatives, provided under and subject to the Gemma Terms of Use and the Gemma Prohibited Use Policy. Hy-MT2 is distributed under the Apache License, Version 2.0.

System integration

• Services: "Translate with Interpreter" appears in the Services menu of any app. On selected text it opens the translator with that text loaded; on a document (for example, right-clicking a file in the Finder) it opens the document window for that file. The item may need one app launch, or a re-login, before macOS lists it.
• Drag and drop: drop a document on the app or its icon to open it for translation. If several files are dropped, the first one is used.
• Open With: Interpreter is offered for any file, and tells you plainly when it cannot read one. File > Open's panel and the Finder service are filtered to the types listed above.

Privacy

Everything Interpreter does with your text and your documents happens on this Mac:

• Translation runs locally, on models you downloaded once.
• Document conversion uses textutil and the bundled pdfutil, both local tools.
• OCR uses Apple's Vision framework, which recognizes text on-device. You can verify this by running the extraction under a network-denying sandbox - it works unchanged.

The app reaches the network for only two things: listing model sizes in the chooser, and downloading a model you asked for.

One honest caveat about OCR: Vision's per-language recognizers are macOS assets. Common languages ship with the system, but the first use of a less common recognition language on a fresh machine can trigger a one-time asset download from Apple. That is a model coming down, never your document going up. If you are offline and the asset is missing, that page falls back to the recognizers already installed.

Requirements

• An Apple Silicon Mac. The app and both of its inference engines are arm64-only; Intel Macs are not supported.
• macOS 14.6 or later.
• Free disk space and memory for at least one model. The smallest model on offer is about 2 GB on disk; the largest is close to 29 GB, and the chooser only shows it to Macs with the memory to run it.
• A network connection to download models.

Installation

Drag Interpreter.app to your Applications folder and launch it. The app is signed with a Developer ID certificate, notarized by Apple, and the notarization ticket is stapled, so Gatekeeper clears it on first launch without contacting Apple. You will still see the usual "downloaded from the Internet" confirmation the first time you open it.

The app is about 143 MB on disk. Models are downloaded separately, from inside the app.

Known limitations

• Automatic source-language detection is not in this release. You pick the From language yourself. It is planned for a later version.
• Two windows can load two models. Each window owns its own engine instance. New windows default to the model you last used, so this usually costs nothing, but explicitly picking different models in two open windows will hold both in memory at once.
• PDF, right-to-left scripts: Arabic and Hebrew text is extracted in visual rather than logical order, so translating from an RTL PDF may be garbled. This is a PDFKit limitation.
• PDF line breaks fall at each visual line, not at each paragraph, so translated output from a PDF arrives hard-wrapped.
• PDF pages that mix a text layer with text trapped inside an image contribute only their text layer. The image text is not recognized on such pages, because the page is not text-less.
• Legacy plain text without a byte-order mark - an ISO-8859-1 or Windows-1252 file, for example - may be mis-decoded. This is a textutil limitation, and automatic charset guessing proved too unreliable to trust. Save such files as UTF-8 first.
• Hy-MT2 and European languages: quality testing found occasional grammar and word-choice slips at 7B, and real mistranslations on European pairs at the 1.8B size. Use it for Chinese and Japanese, and prefer TranslateGemma or MiLMMT-46 for European languages.
• Whitespace inside a translated chunk may be normalized by the model. Separators between chunks are preserved exactly.

Credits

Interpreter is built with:

• OMC, a low-code app builder for macOS
• ActionUI, a JSON to SwiftUI interface builder
• mlx-agent, a native Swift local-LLM agent built on MLX Swift (Apache 2.0)
• MLX and MLX Swift, Apple's machine-learning array framework for Apple Silicon (MIT)
• llama.cpp by Georgi Gerganov and the ggml contributors, which serves the GGUF models (MIT)
• pdfutil, a PDFKit-based command-line PDF toolbox (Apache 2.0)

Interpreter itself is licensed under the Apache License, Version 2.0. Model licenses are listed per model in the app's info sheet and in the NOTICE.txt installed beside each model.

The full license text for every bundled component ships inside the app, under Contents/Support: MLX/mlx-agent.LICENSE and MLX/mlx-agent.THIRD-PARTY-NOTICES.txt (which covers the Swift packages linked into the translation engine), pdfutil.LICENSE, and Llama.cpp/LICENSE.

---

This discussion was created from the release Interpreter.app 1.0.