Distillation

Distil a model into a file that runs on a CPU.

Eldric can compress what a model knows into a compact, self-contained file that runs on an ordinary CPU — no GPU, no Python at runtime, small enough for a Raspberry Pi. The knowledge becomes an artifact you can ship.

01 — The idea

Knowledge, compiled into a file.

Instead of carrying a large model everywhere, distil what it knows for a specific job into a compact file. The result loads fast, runs on CPU, and is small enough to deploy at the edge — a model's competence on a narrow task, packed into kilobytes.

02 — Native fold

Structured-ML runs natively — no separate runtime.

The distilled workloads run inside the main Eldric service: time-series forecasting, control policies, and associative retrieval execute natively, with no separate xLSTM daemon. One package to install and update.

03 — What you can distil

From a teacher to a compact file.

  • A control policy for a named mechanism — drive it toward a target from the current state.
  • The dynamics of a modelled physical system — roll it forward step by step on CPU.
  • A time-series forecaster — extend a recent window of a signal.

Each is a forward model of a known system, downloadable and CPU-class. Download the compact demo models →

04 — Honest scope

What it is — and isn't.

Forward models, bounded windows. A distilled model rolls a known system forward; it does not predict a real-world event, and it is faithful over a bounded window, then drifts — the nature of the method, stated plainly. We distil competence on named tasks, not general intelligence.
Get started

Try it.

Install the platform — dnf install eldric-aios on RHEL / Fedora — and load a compact model. See Get started or what Eldric does.