Hardware tiers
Three tiers, each with an honest user expectation. The preflight-check measures
your hardware and recommends one; see performance.
| Tier | Experience | Models | USB | RAM | Engine note |
|---|---|---|---|---|---|
| ”runs almost anywhere, quick help” | 1× 4-8B + local role council + Whisper-base | 16-32 GB | 8-12 GB | CPU ok | |
| Field | ”documents, notes, stable offline” | 8B + Vision + Whisper | 32 GB | 12-16 GB | GPU recommended |
| Lab | ”real local multi-model council, deep” | 32B reasoning + 32B agent + 8B + Vision | 64-128 GB NVMe USB-SSD (USB 3.2 Gen 2; Gen 2×2 preferred) | 48-64 GB | GPU required |
Correction (council consensus): the v0.1 “Council” preset claimed 24-32 GB RAM. That was unrealistic for two resident 32B models (~40 GB of weights alone, plus KV-cache). Correct Lab tier: 48-64 GB RAM + GPU + NVMe USB-SSD.
On anything below the Lab tier, use the local role council rather than the local multi-model council.