Language Model

The model is the engine behind every reply: it reads your message and writes the answer. The most consequential choice you make in Loca is where that model runs. A Provider is the source of a model, and Loca sorts providers into two kinds, Local and External, which you choose between under Language Model in Settings.

Local or External

A Local model runs on your own machine, through your Core. Nothing about the conversation leaves the device, which is the whole point of running locally, and the cost is that speed and capability are bounded by your hardware. An External model lives behind a provider you configure over the network. It can be faster and more capable than anything your machine would run, in exchange for sending your messages to that provider to be answered.

	Local	External
Where it runs	Your machine, via the Core	A provider over the network
Your data	Stays on the device	Sent to the provider
Speed	Bounded by your hardware	Often faster
Setup	A one-time model download	Provider URL and key

Smart presets

You do not have to read benchmark tables to pick a Local model. A smart preset chooses one that fits your machine: you tell Loca the tier you want, and it sizes the download to your hardware so the model runs well rather than thrashing. This is the path most people should take to a working Local setup.

A model is several gigabytes, so Loca always asks before it downloads one, and never fetches in the background. Once a model is on disk it is shared across channels, so you pay that download only once even if you run more than one build of loca.

Context and speed

Two settings shape how a Local model performs, and they trade against your machine’s RAM and CPU:

• Context Window is how many tokens the model can consider at once. A larger window holds more of the conversation, which matters for long sessions, but uses more RAM. This is the same window the session bar tracks as you talk.
• CPU Threads is how many of your processor’s cores the model runs on. More threads can speed up replies on a capable machine, up to the point where you run out of cores to give it.

Loca can also unload an idle model after a while, so a Local model does not hold gigabytes of RAM when you have stepped away. It loads again on your next message.

Thinking and Creativity

Two dials change how the model answers from one message to the next. You reach both from the quick controls that sit with the input bar, so you can adjust them mid-conversation without opening Settings. They are per-message choices, not a mode you commit to.

Thinking is how much the model reasons before it replies. Turn it off for a fast, direct answer to something simple; raise it when a question needs the model to work through steps or check itself, the kind of problem where a snap answer tends to be wrong. More thinking is not free: it takes longer and spends more of the context window, since the reasoning itself occupies the model’s attention. It shows up as a Thinking bubble above the reply, which you can expand to read.

Creativity controls how freely the model chooses its words. Keep it low when you want factual, repeatable answers; raise it for brainstorming or varied phrasing. The two dials are independent and often pull in opposite directions: a careful factual answer wants thinking up and creativity down, while a freewheeling idea session wants the reverse. Both are requests to the model, not guarantees, and two models may respond to the same setting differently.

How It Works

A Local provider runs the model as a service on the Core’s machine: the Core loads the model file into RAM (and onto the GPU when one is available) and answers each turn through a local inference server, so a Local model’s speed comes straight off your own hardware. An External provider skips all of that and forwards the turn to the API you configured. Either way the Core speaks the same request shape, which is why switching providers does not change anything else about how Loca works.

A smart preset leans on a hardware-fit step: at install Loca reads your machine’s RAM, processor, and graphics memory, then picks the largest model and quality level that should run comfortably in that budget, rather than handing you one that will swap to disk. The dials in this chapter ride along with each turn: Context Window and CPU Threads are set when the model loads, while Thinking and Creativity are sampling instructions sent with the message itself, which is why they can change from one reply to the next.

With where the model runs settled, you can pair it with a second one for the hard questions: Dual Model.