HomeProjectsAmberlin Settings
ToolGTK4CPU · CUDALinux Mint

Amberlin Settings

Which model Amberlin runs, and whether this machine can run it. One GTK4 window over the model archive: the precision each export ships at, the runtime it needs, the video memory it wants and the bytes still to fetch — with the models already on disk marked.

Why a tool for this

Picking a local model looks like picking a download and is not. The 4-bit export of the model Amber ships is larger than the 8-bit one; an export that exists is not always an export that loads; and a file built for a runtime you are not using can start, run, and answer with an exclamation mark repeated until it runs out of room. Every one of those was found by measuring rather than reasoning, and the measurements are written up here.

This is the tool that puts that reading in front of the machine it applies to. It is not a settings panel Amberlin happens to have — it is a separate package with one job, which is why the archive files it under tools rather than applications.

Quick start

bash
# Install
sudo apt install amberlin-settings

# The runtime it reports against — one of these, never both
sudo apt install amberlin-runtime        # CPU: works on any machine
sudo apt install amberlin-runtime-cuda   # NVIDIA, Blackwell included

# The weights it sizes up: the ear, the voice, and the answer
sudo apt install amber-models

What it is for

One choice for the whole desktop

The models are hosted per-user by amberlin-backend on the session bus, not inside any one window. The model chosen here is the model every application asking gets — Amberlin, kat800, and anything else that makes the call.

Size is not the same as cost

Exports are listed with the precision they ship at rather than sorted by megabytes. For the model Amber ships, the 8-bit export is both smaller and more precise on the weights that do the reasoning than the 4-bit one — the download column alone would recommend the wrong file.

The runtime decides what is worth offering

CPU or CUDA is a different package, and only one of them is installed. An export is chosen for the runtime that will open it, so what this machine can actually load is part of the same view as what it could download.

A preview, and a destination

Amber ships a small model that answers immediately on any machine and treats it as a preview rather than the destination; the larger one is worth its download only with a GPU under it. Which side of that line a machine falls on is exactly what this window answers.

Three modalities, three decisions

The ear, the voice and the answer are separate packages, and the language model is the big one. A machine that only wants dictation can leave it out — and this is where the size of that decision is visible before it is taken.

Estimated says estimated

A memory figure worked out from the shape a model declares is marked apart from one measured on this machine while it ran. Both are useful and they are not the same promise, so the tool never quietly presents one as the other.

What it shows per model

Six figures, and each one decides something. A model is worth its download when they agree; when they do not, the disagreement is the answer.

ShownWhat it decides
QuantisationWhich export — full precision, half, 8-bit, 4-bit. Precision and file size do not order the same way, so both are shown.
RuntimeCPU or CUDA, as installed. The pairing matters: a file built for the wrong runtime can load and still answer with nonsense.
Video memoryThe weights, plus the cache the model keeps for the conversation. That cache cost is a property of the model’s geometry, so a budget that is generous for one is punitive for the next.
DownloadThe bytes still to fetch before this model can answer anything.
On diskModels already fetched are marked, so changing your mind carries a visible price rather than a surprise one.
ProvenanceWhether the memory figure beside a model was estimated from its declared shape or measured while it ran.

The weights themselves are ordinary packages — one per modality, so a machine that only wants dictation can leave the language model and its download out. They are listed with the rest of the archive on the packages page.