A supervised process that owns a loaded Magika instance.

The server loads the ONNX model and configuration once in init/1 and publishes the resulting Magika.t() into :persistent_term. The model and configuration are immutable and ONNX Runtime inference is thread-safe, so callers fetch the instance with instance/1 and run inference directly in their own process — there is no serialization through the server's mailbox and no per-call copying of the (large) configuration.

The server therefore owns the lifecycle of the instance: it loads it on start and erases the published term on shutdown. It is normally started for you by Magika.Application, but you can also place it in your own supervision tree:

children = [
  {Magika.Server, prediction_mode: :high_confidence}
]

Multiple independently-named servers can coexist:

{Magika.Server, name: MyApp.Magika, model_path: "..."}

and are then addressed via Magika.identify(content, server: MyApp.Magika).