A knowledge library stores inspectable sources. Model weights are a separate learned component. More stored information does not automatically make the language model capable.
Storage planning
These are scenario assumptions, not a hardware purchase or a promise of model intelligence. Working storage includes raw sources, extracted text, index allowance, training corpus, scratch space, three checkpoints and 25% free headroom.
Backups are separate. RAM/VRAM values exclude activations, caches, runtime and operating-system overhead. Original training code currently uses CPU Float64 arrays; lower-precision/GPU figures describe a future implementation, not a supported runtime.
References: training memory components · model/data/compute scaling research.