Meet your next stack.
Before you download it.

What the result actually means

This is a conservative planning estimate for one active user at roughly 4k context, not a performance guarantee. It budgets separately for application services, embedding, operating-system headroom, model weights, and inference context.

GPU mode requires the entire model budget to fit dedicated VRAM. CPU mode uses system memory; Apple mode uses a single unified pool without adding it twice. Larger documents, simultaneous users and longer context can need substantially more memory.

Remote APIs remove local chat-model weight requirements. They do not remove application memory requirements, token costs or the need for an internet connection. Every suggested model still needs compatibility verification in your chosen backend.