Advanced

How Primordial manages AI models.

Learn when models are downloaded, loaded, shared, protected during active work, and removed from memory.

Activation is required

Installing or running models requires accepted evaluation activation. See Evaluation Keys to activate Primordial for development and device testing.

Downloading and loading are separate

Downloaded means verified model files are installed for offline use. Loaded means the model is in memory and ready to run. Installation never loads a model, and loading never downloads one.

Models are shared across clients

Copies of PrimordialClient share one runtime. When they use the same model, Primordial shares its model weights and coordinates access instead of loading duplicate copies.

Active work protects models

Primordial prevents a model from being unloaded, replaced, or removed while an operation is using it. Loading a model through the advanced API keeps it in memory until you unload it or Primordial responds to system pressure.

Some models cannot run together

Workflows that use your data prepare embeddings before starting generation. Generation and speech models do not remain in memory together. If active work prevents Primordial from switching models, the operation returns resourceConflict.

Primordial frees memory automatically

When your app enters the background, the Mac sleeps, or the system reports memory pressure, Primordial unloads models that are not being used. Active work completes or cancels normally and is never removed from memory unexpectedly.