Localize is a private-first desktop workspace for translating text, documents, and images with models that run on your own computer. It is built for the moments when copying sensitive text into a web translator is not an option—or when you simply want to choose the model that does the work.
- Translates text, detects its language, and lets you replace a translated word or phrase with contextual alternatives without translating the whole text again.
- Extracts and translates text from PDF, EPUB, MOBI, DOCX, XLSX, and PPTX files. Scanned pages can fall back to local vision OCR.
- Translates text in images and extracts text through a separate OCR workspace.
- Supports Ollama as well as a Localize-managed llama.cpp server, so models can come from Ollama or public Hugging Face GGUF repositories.
- Adds local speech input through whisper.cpp: record from a selected microphone, transcribe an audio file, or insert live speech into the editor.
- Keeps translation content out of a history database. Temporary input files, rendered pages, and recordings are cleaned up after the active operation.
Localize is a Windows 10/11 x64 application. Native runtimes and model weights are deliberately not shipped with the installer: choose exactly what you need from Settings after the app is installed.
Download the latest Windows x64 executable or NSIS installer from the GitHub Releases page. After launching Localize, choose and install the inference runtimes and models you want in Settings; they are not included with the application download.
To develop Localize yourself, install Go 1.26+, Bun, and a Windows WebView2 runtime. Then clone and start the app:
git clone https://github.com/FlameInTheDark/localize-app.git
cd localize-app
bun --cwd frontend install
go run github.com/wailsapp/wails/v2/cmd/wails@v2.12.0 devFor the quickest start, install and run Ollama, then
open Settings → Models in Localize and pull a translation model such as
translategemma:latest. Pull a vision-capable model separately if you want to
use image translation or OCR.
If you prefer GGUF models, open Settings → Runtime to install a selected
llama.cpp CPU, NVIDIA CUDA, Vulkan, or AMD HIP/Radeon build. Choose the runtime
type that matches your hardware; automatic mode continues to use CUDA when an
NVIDIA runtime is available, otherwise CPU. In Settings → Models, choose
where GGUF files should live, find a public Hugging Face model, download it,
and assign it as the translation model. Vision models also need their compatible
mmproj projection file.
You can switch between Ollama and llama.cpp at any time. Localize starts only the llama.cpp processes it owns, listens on loopback, and stops them when the app closes.
These are the models currently used during Localize development. They make a good local-first starting point and are not included with the app download.
| Use | Ollama | llama.cpp / whisper.cpp |
|---|---|---|
| Text translation | translategemma:latest (3.3 GB) |
translategemma-4b-it.Q8_0.gguf (3.8 GiB) |
| Image and document OCR | glm-ocr:latest (2.2 GB) |
GLM-OCR-Q8_0.gguf with its matching mmproj file (about 1.3 GiB together) |
| Speech input | — | ggml-large-v3-turbo.bin (1.5 GiB) |
For llama.cpp vision/OCR, always download the model and its matching mmproj
projection file. If your hardware needs a smaller model, choose another public
GGUF or Whisper model in Settings and assign it to the same role.
Voice input is optional. Install a chosen whisper.cpp version under Settings → Runtime, then select an official Whisper model and microphone in Settings → Speech. The Voice button records or accepts WAV, MP3, FLAC, and OGG files; Live inserts completed speech into the current text editor.
Document translation needs MuPDF. Pick and install an official MuPDF Windows version from Settings → Runtime before translating documents. Localize returns aligned source and translated text; it does not rebuild the original PDF or Office layout.
Configured inference stays local: Ollama is contacted on its local endpoint,
and llama.cpp runs on 127.0.0.1. Localize only uses the network when you ask
it to download a runtime or model, or when you choose a provider that uses one.
Runtime and model downloads show progress, speed, and remaining time, are
resumable where upstream supports it, and are published atomically after
validation.
Run the checks used by the Windows verification workflow:
go vet ./...
go test -race ./...
bun --cwd frontend run check
bun --cwd frontend run buildBuild a production executable with:
go run github.com/wailsapp/wails/v2/cmd/wails@v2.12.0 buildTo build the NSIS installer instead:
.\scripts\build-installer.ps1Pull requests and pushes to main run the Windows frontend and Go checks.
Conventional commits on main create semantic releases: feat: creates a
minor release, fix: creates a patch release, and a breaking change creates a
major release. The release workflow attaches both the Windows executable and
the NSIS installer. Runtimes and models remain user-installed rather than
release assets.
The architecture is documented in docs/ARCHITECTURE.md. See LICENSE for Localize's license and licenses/THIRD_PARTY_NOTICES.md for the licenses and sources of third-party components.