2. Core training.

Private
corpus
Base
weights

Domain depth at volume.

Train across large volumes of unstructured internal material so the system absorbs your vocabulary, formats and edge conventions.

Architecture choice.

Pick between dense and sparse-routing layouts to balance throughput, specialisation and cost against the workloads you actually run.

Mixed-input ready foundations.

Text, image and audio inputs are carried through the same pipeline wherever your workflow calls for them.