Skip to content
MetalGlot Docs

MetalGlot Docs

MetalGlot is built on a local-first translation stack. These docs cover both everyday app workflows and the technical pieces behind them, so you can move from first-run setup to backend details without switching documentation sets.

Product Guides

Section titled “Product Guides”
Getting Started Set up a container, download a model, and get to your first translation quickly.
Interface Language Selector Switch the app UI across 32 localized interface languages.
Translation Pages and Workspaces Choose the right workspace for JSON, XLIFF, ARB, Markdown, subtitles, raw text, and image tasks.
Translation Wizard and Validation Warnings Review translations one item at a time and catch format-breaking issues before export.
GGUF Models and Downloads Download, load, pause, resume, and manage the models your container uses.
Inference Settings Tune generation controls and decide when the English bridge is worth using.
Docker Images and Containers Choose the right image, create a container, and manage model folders and resources.

Platform Docs

Section titled “Platform Docs”
FastAPI Service Understand the local API layer that powers translation and model lifecycle operations.
Docker Images Review the backend image variants used for CPU and CUDA deployments.
Models and Quants Compare TranslateGemma model families, GGUF quants, and their storage tradeoffs.
Fine-Tuning TranslateGemma Follow the Welsh reference workflow for adapting TranslateGemma to a new language.

Why This Matters

Section titled “Why This Matters”

MetalGlot is designed for users who want translation workflows that stay on their own hardware. That means the implementation details matter:

  • The interface language the operator is most comfortable working in.
  • The workspace that is best for the source format you are translating.
  • The container and model folder that are active on the current machine.
  • The API layer that is running locally.
  • The image that is best for the machine you are deploying on.
  • The model size and quant level you should choose for your hardware budget.
  • The way to adapt TranslateGemma to a new language without losing instruction-following behavior.

Source Projects

Section titled “Source Projects”

Model Credits

Section titled “Model Credits”

MetalGlot uses the open TranslateGemma model family published by Google and GGUF quantizations published by mradermacher.

Each model family is available in multiple quant levels inside MetalGlot, not just a single GGUF file. The goal is to let users trade memory use, speed, and output quality based on the machine they are actually running.