The Model Control Pill
Every button-driven engine or model tier in Admin → Model Manager — Coqui, and Qwen's Base and higher-quality tiers — gets the same pill-and-button control: a status pill on the left (idle / loading / ready / unreachable) and a Load or Stop action on the right. It's the one place the app tells you what's actually resident in GPU memory right now, separate from what's merely installed on disk.
Below, Kokoro's row in a ready state: the pill reads a green "Kokoro v1 ready," with DEFAULT, FALLBACK, and verified badges, and the action is Stop — the same control that loaded it now unloads it and frees the VRAM.
An engine still loading shows an amber "Loading …" pill with its button
disabled, so a second click can't fire mid-load; one that's installed but not
yet loaded reads "idle" with a Load model action instead. An engine
without an installed package yet (e.g. Coqui before its pip install step)
or without its weights fetched (e.g. Kokoro before its one-time download)
shows an install/repair action instead of this pill — there's nothing to
load until that step completes. See Voice Engines for what
each engine's row looks like in those states.
Below, llama3.1:8b: installed on disk but not resident in GPU memory, so the pill reads a neutral "llama3.1:8b idle" and the action is Load model rather than Stop.
Click Load model and that same row flips to an amber "Loading llama3.1:8b…" pill with its button disabled and relabeled Loading…, so a second click can't fire mid-load.
Next: Library Management.