Skip to content

Model Manager

dudarenok-maker edited this page Jul 5, 2026 · 1 revision

Model Manager

Model Manager (#/models, reached via the Open Model Manager button on Admin) is where every local model — TTS engines, the ASR model, and the local analyzer — gets installed, updated, removed, and loaded or unloaded from the GPU, all from one place. It's also where the settings that used to live on the Account page now live: default engine per model kind, the two-model analyzer split, TTS sidecar tuning, and server configuration.

Device

Device panel — detected GPU and VRAM

The Device panel shows the GPU(s) detected on this machine and their VRAM — this is what determines which engines are even offered (a CPU-only box still runs Kokoro; a GPU box can run Qwen/Coqui too).

Installed models

Installed models — Standard group

Every model is grouped under Standard (Kokoro, Qwen3-TTS Base 0.6B/1.7B, Qwen3-TTS VoiceDesign), Optional add-ons (Coqui XTTS v2), Analyzer models (Ollama), and Speech recognition (ASR). Each row shows:

  • Disk size + path — read straight off the model files on disk, not a cached estimate.
  • Status badgesNot installed, Weights missing / Needs repair (present but broken — offers a Repair action instead of Load), Installed, or Loaded, plus Default / Fallback tags for the engines that hold those roles, and an integrity chip (verified / mismatch / unpinned) for fixed-file models.
  • Load / Stop — the same Model Control Pill used everywhere else in the app, when the row has a usable install.
  • Install / Update / Repair — expands an inline installer under the row for engines that ship one (Kokoro, Coqui, Qwen Base, Whisper).
  • Remove — deletes the model's weights from disk, gated by a confirm dialog that explains and blocks the three cases the server itself refuses: the model is currently loaded, it's the universal fallback engine, or it's your current default engine.

Defaults for new books

Defaults for new books

Knob What it does Default Range
Analysis model Model used to analyze new books account.defaultAnalysisModel curated list ∪ live Ollama tags
Voice engine Default TTS engine for new books account.defaultTtsEngine Kokoro / Qwen3-TTS / Coqui
Voice model Default voice model, scoped to chosen engine resolved defaultTtsModelKey engine-dependent

Two-model analyzer split (advanced)

Two-model analyzer split

Splits analysis across two models so cast detection and sentence attribution run concurrently on different free-tier rate-limit buckets. Off by default — both phases use your default analysis model until you opt in.

Knob What it does Default Range
Phase 0 model (cast detection) Model used for cast detection blank = server default dropdown
Phase 1 model (attribution) Model used for sentence attribution blank = server default dropdown
Phase 1 minimum chapter lag Chapters Phase 0 must clear before Phase 1 starts 10 (effective) integer, 0–50

Voice engine

Voice engine settings

Knob What it does Default Range
Auto-start with server Starts the TTS sidecar automatically with the server (needs server restart) true boolean
Keep both voice engines loaded Dual-model mode — both engines resident at once false boolean
Generation workers Chapters synthesized concurrently 1 integer, 1–4

Server configuration

Server configuration

Knob What it does Default Range
Voice engine URL Base URL of the TTS sidecar http://localhost:9000 string, private/loopback host only
Analyzer engine Routes analysis through Gemini direct vs. local Ollama for THIS account gemini gemini / local
Ollama URL Base URL of the local Ollama daemon http://localhost:11434 string
Gemini API key API key for Gemini analyzer/persona calls (unset) string

Not the same knob as Advanced Settings's "Analyzer engine." This one is your per-account preference (defaults to gemini, matching the free-tier-friendly out-of-the-box experience); Advanced Settings' analyzer.engine is the lower-level server/env config knob (defaults to local), used only when no per-account preference is set. Same English label, two different controls.

Install / update analyzer (Ollama)

Install / update analyzer

Install the Ollama daemon and pull analyzer model weights without dropping to a terminal — pull-tag UI plus a live health probe. The voice-engine / ASR models (Kokoro, Qwen, Coqui, Whisper) install from their own rows in Installed models above, not here.

Next: Advanced Settings.

Clone this wiki locally