local-model-opslisted
Install: claude install-skill SenjuWoo/Ultimate-AI-Starter-Bundle
# Local model ops
Wiring a local GGUF served by LM Studio into Hermes. Runtime-dependent details
come from inspected installations; recheck them after upgrades. Historical
measurements are examples, not guarantees for another model or machine.
## Why run local at all
Choose local or hosted by measured quality, speed, budget and privacy needs.
Local inference can provide:
| Reason | What it gets you |
|---|---|
| **Model choice** | Choose models suited to adult, red-team or dark-fiction work; behavior still depends on the model and harness |
| **Local inference** | Prompts stay local only when auxiliaries, fallback routes and tools are also local or disabled |
| **No hosted inference quota** | Offline inference after setup; local servers can still reject overloaded requests |
Compare hosted options in `token-efficiency`; do not assume either route is cheaper or faster.
## Hermes speaks LM Studio natively
`lmstudio` is a **first-class provider**, not a generic OpenAI-compatible
endpoint. That buys real behaviour you do not get from a custom provider:
- Default route `http://127.0.0.1:1234/v1`, overridable with `LM_BASE_URL`.
- **No credential setup.** Hermes supplies its own placeholder key. Never put
a real secret in an `lmstudio` provider block; there is nothing to
authenticate to.
- Explicit preload: Hermes loads the model into LM Studio itself before the
first turn. Set `model.lmstudio_load_mode: jit` to skip that and let LM
Studio load on first request inst