github tanakh/plamo-translate-desktop v0.4.0
PLaMo Translate Desktop 0.4.0

latest releases: v0.6.0, v0.5.0
14 days ago

Added

  • macOS (Apple Silicon) support. scripts/package-macos.sh builds the package: the app in a disk image, for macOS 13.3 or later. llama.cpp runs on the GPU through Metal. The Releases page has it as …-macos-arm64.dmg. It is not notarized: the README says how to open it.
    • Connect registers the browser host with Chrome, Chromium, Edge, and Brave in ~/Library/Application Support. It used to write to ~/.config, where only their Linux versions look.
    • The GPU is used on macOS 14 and 15 as well, where llama.cpp 0.8.0's Metal shaders do not compile with BF16 enabled (the PLaMo models do not use BF16).
    • The app has a menu with Quit (Cmd+Q) and quits when its window is closed. Quitting while a model is loaded, or while one is loading, exits cleanly, in the app and in the browser host.
    • Models on the GPU stay memory-mapped, since Metal uses the mapped file directly: with 0.8B, a peak of about 200 MB of memory instead of 760 MB, at the same speed.
  • The browser extension translates on its own too, with the model running in the browser: wllama (llama.cpp compiled to WebAssembly) on the CPU or the GPU (WebGPU). The app is not needed for it. See "Without the app" in the README.
    • Engine under Settings in the panel: Automatic (the app when it is connected, otherwise the browser), PLaMo Translate Desktop, or In the browser.
    • The 0.8B and 2B models (Q4_K_M), downloaded into the browser from the panel after accepting the PLaMo Community License (resumed after a network error, and deletable). The larger ones do not fit in WebAssembly's memory.
    • Device: Automatic measures the CPU and the GPU on the first translation and keeps the faster one; it asks for the high-performance GPU (the README explains Windows' GPU assignment). With 2B, generation runs at about 48 tokens/s on a Ryzen AI Max+ 395 (Radeon 8060S), 67 on a Radeon RX 9070 XT, and 18 on the CPU of a Ryzen 9 9950X.
    • Released after 5 minutes without use, or with Release, like the app's host.
    • The panel links to the README's steps for setting up the app, which the larger models need.
    • A package with the extension alone (…-browser-extension.zip), for trying it without the app.
    • The status shows the device really used, such as WebGPU (amd rdna-4), and says so when the GPU could not be used. With the device set to the CPU, the number of threads can be chosen under Settings.
    • Page translation runs four batches of paragraphs side by side, dealt in turn so that the page fills from the top: with 2B on a Radeon RX 9070 XT, 80 paragraphs take 47 s instead of 70. Each paragraph then has less of the text just before it as context; Translate pages in parallel under Settings turns it off.
  • The panel shows the speed in tokens per second while it translates and when it is done, and the box on a translated page shows the page's speed.
  • While the page in the current tab is being translated, the panel's page button stops it.
  • A pulsing frame on the page marks the paragraphs being translated (with the in-browser engine, up to four at once). Settings in the panel can turn it off.
  • Page translation skips paragraphs that do not show (collapsed sections, menus, hidden text) until they do. Settings in the panel can also limit it to the paragraphs near the view, translating the rest as you scroll to them.
  • The panel shows what Automatic uses: the engine (the app or the browser) and the device, such as Automatic (GPU) with the speeds measured in the browser, or the device the app measured. Measure and choose under the device measures the browser's GPU and CPU again.
  • Reset settings under Settings in the panel: back to a first start (the license acceptance, the choices, the measured device); downloaded models are kept.

Changed

  • The extension needs Chrome or Edge 137 or later.

Fixed

  • Downloads in the app: PFN replaced every model file on 2026-11-06 (only the chat template changed, which the app does not use), and the files of the first revision are no longer served. The built-in model list now points to the new revision, and models downloaded before still count as downloaded.
    • Partial downloads of the first revision, which can no longer be resumed, are removed (they could be gigabytes, and were neither shown nor deleted).
  • Output that repeats itself is retried with the repetition penalty (DRY) as designed: a setting of llama.cpp's had left it off, so the last retry ran the same way as the first.
  • With Auto off, a model that finishes loading no longer starts a translation, unless one was asked for while it loaded.
  • A download finishing no longer replaces a model chosen while it ran; the first download to finish becomes active only when no model is.
  • Swapping with Japanese ⇄ English as the target and an explicit source language sets the source to the translation's language.
  • Unloading or deleting a model while it loads no longer leaves "Loading… 100%" in the status bar, and a model whose files cannot be deleted stays selected.
  • A settings.json saved with a byte order mark is read, and one in another encoding than UTF-8 is kept aside instead of being overwritten.
  • The browser extension: closing the panel while it translated no longer stopped a page translation in progress; a page translation that fails stops all its batches; translations of a stopped page translation no longer reach the next page in the tab; and only the extension's own pages can connect as its panel.

What's Changed

  • Package the app for macOS on Apple Silicon by @kozo2 in #1

New Contributors

  • @kozo2 made their first contribution in #1
  • @fronori contributed the macOS fixes in #2 (the GPU on macOS 14 and 15, and quitting with a model loaded)

Full Changelog: v0.3.0...v0.4.0

Don't miss a new plamo-translate-desktop release

NewReleases is sending notifications on new releases.