hyperquader

Amberlin is the speech-first chat surface of Amber Linux: hold a key, ask out loud, and the answer comes back spoken and drawn. A local tool server exposes the machine — files, git, the clock, the weather, system readings, a note store — and the assistant gets things done by calling those tools.

This organisation publishes the models behind it, in the shape the desktop actually loads: ONNX Runtime GenAI model directories, which run without PyTorch on a CPU or a GPU. Amber Linux fetches them by commit and verifies every file by hash, so a build here is something a machine downloads unattended rather than something a person evaluates.

Two kinds of artefact, and the difference is stated on every repo:

• A conversion. An upstream model, converted and quantised, with nothing learned added. The weights are the publisher's; the format is ours.
• A fine-tune. A LoRA adapter trained on the amberlin-tools call format, published as the adapter, the merged weights and a converted build.

Every claim is measured on a held-out split and reported with the numbers that produced it — including the ones that did not come out well. A tuned model is only published when it beats the untuned one on the same split, at the precision that ships, against the same tool server. Until it does, what is here is a conversion, and the card says so.

https://hyperquader.com · https://amberlinux.org