Skip to content

Docker Images and Containers

MetalGlot runs its local AI translation backend in Docker. Before you can translate anything, you need to download the right CPU or CUDA image, create a container from that image, and start it from the app.

Illustrative MetalGlot Docker setup workflow showing available images, pull progress, and container creation settings

Open Setup Container from the home page toolbar to see the available image variants.

VariantBest forApproximate compressed download size
CPU onlySystems without a compatible NVIDIA GPU1013 MB
CUDA Legacy 10xx PascalNVIDIA 10-series GPUs4.31 GB
CUDA Mainstream 20xx/30xx/40xx/A100Most current NVIDIA desktop GPUs and A100-class systems4.42 GB
CUDA Next-gen 50xx/B200Newer NVIDIA 50-series and next-generation compatible systems7.12 GB

If you are not sure whether a CUDA option matches your GPU generation, use the CPU image instead of forcing the wrong CUDA variant.

If you want the simplest and safest starting point, begin with the CPU image. Choose a CUDA image when you know your NVIDIA generation matches the listed variant.

  1. Open Setup Container.
  2. Review the available image variants.
  3. Choose Download for the variant that matches your hardware.
  4. Watch the Progress monitor until the download completes.

Docker image downloads and model downloads are separate. You download the container image first, then you download translation models later from the model manager.

Once an image is present, choose Create container for that image.

During container creation, MetalGlot asks you to configure the model folder and optional resource settings.

This is the most important container setting.

The folder you choose becomes that container’s model library from the app’s point of view. When you later open Manage AI models, MetalGlot shows and uses the folder that belongs to the currently selected container.

That folder is where the GGUF models and vision models for that container are stored.

That means every container can have its own download folder.

Container purposeExample host folder
CPU setupD:\MetalGlot\models-cpu
GPU setupD:\MetalGlot\models-gpu
Experimental setupD:\MetalGlot\models-test

Using separate folders keeps lightweight CPU downloads separate from larger GPU-focused model libraries.

MetalGlot offers two CPU modes:

  • Use all CPU resources.
  • Specify CPU threads.

Start with the default unless you have a reason to limit CPU usage on a shared workstation.

CUDA containers also expose GPU controls:

  • Use entire GPU resources.
  • Advanced: specify GPU layers.

When advanced GPU tuning is available, MetalGlot shows detected GPU memory and an estimated layer recommendation for the currently selected model family. Use that estimate as a starting point, not as a strict rule.

Starting, Stopping, And Switching Containers

Section titled “Starting, Stopping, And Switching Containers”

Back on the home page toolbar, you can:

  • Select a container from the container picker.
  • Start the selected container.
  • Stop the running container.
  • Delete a container you no longer need.

MetalGlot expects one active translation container at a time. If another MetalGlot container is already running, stop it before starting a different one.

How Container Selection Affects Model Management

Section titled “How Container Selection Affects Model Management”

The selected container controls:

  • Which model folder is shown in Manage AI models.
  • Which downloaded models are visible from the app.
  • Which local backend is currently serving translation requests.

If you switch containers, you are also switching to that container’s model folder and available model library.

Docker image downloads show progress in the Docker setup dialog.

  • You can cancel an in-progress image download.
  • After cancellation, you can restart the image download later.
  • Docker image downloads do not use the same pause and resume workflow as GGUF model downloads.
  • If MetalGlot says Docker is not detected, install Docker Desktop first.
  • If Docker is detected but not running, start Docker Desktop and wait for it to finish launching.
  • If a container exists but will not start, make sure another MetalGlot container is not already running.
  • If model downloads are unavailable, confirm that the selected container was created with a valid model folder.