Model preloading
When Preload on startup is on, Loach sends an empty chat to your default model as the app launches - warming it into VRAM so the first real prompt of the session streams without the usual cold-start pause.
Enabling
Turn on Preload on startup in Settings → General, right under the Default model picker. It is off by default, and it is Ollama-only - hosted providers do not have a meaningful concept of warm vs cold from your side. While the default resolves to an OpenAI-compatible model (or to no model yet), the switch is greyed out with a note to pin the default to an Ollama model.
The trade-off
Preloading pins VRAM at launch even if you only opened Loach to re-read an old conversation. That is why it ships off: many users only want VRAM consumed when they are about to actually chat. Turn it on if you launch Loach with intent - you know the first thing you do is going to be a fresh prompt.
Which model gets warmed
Preload resolves through the same logic as Default model:
- If you pinned a specific model, that one.
- If you pinned to a provider, the model of your most recently active chat on that provider.
- Otherwise, your most recent (provider, model) pair.
Only an Ollama result is warmed - if the default lands on an OpenAI-compatible model, preload does nothing. A Space’s own default model doesn’t change what gets warmed.
When it runs
- At launch - warming starts while the window is still loading, and is repeated once your settings and chats have loaded. With Auto-launch Ollama on, Loach waits for the server it started before warming.
- After unlocking - with an app lock set, nothing is warmed until you unlock. It happens once per launch: unlocking again later in the session doesn’t warm the model a second time.
- Sized like your first chat - the model is loaded with the context length, GPU layers and Low VRAM setting a new chat with it would use, so your first message doesn’t trigger a reload.
- Silent - if Ollama isn’t running or the model isn’t installed, the preload simply doesn’t happen; no error is shown.
Keeping it loaded
Preloading only helps the first prompt. Whether the model stays resident afterwards is Ollama’s keep-alive: Settings → Features → Keep model loaded sets it to 5 min, 30 min, 1 hour or Always. Loach sends that value with every request - the preload included, so the warmed model doesn’t drop out before you type - and a reply after a pause skips the cold reload too. Until you pick an option, Ollama’s own default applies (5 minutes, unless OLLAMA_KEEP_ALIVE says otherwise). Switching a chat to a different model unloads the previous one regardless.