What the settings do

If you've come from another front end, the first thing you'll notice is that our settings panel is nearly empty. One slider called Imagination, a switch to turn it off, and some controls about recaps. No top-k. No top-p. No repetition penalty, no mirostat, no DRY, no XTC, no sampler order.

That's deliberate, and it's the thing experienced players ask about first, so here's the thinking.

The usual sampling panel is twenty numbers in a scrolling column, most of them named after the maths instead of the effect. It looks like power. Mostly it's a place to fidget.

Here's the problem. Almost none of those values are things you decide per conversation. They're properties of a model and how you're running it. A sensible top-k for one fine-tune is a sensible top-k for every story you'll ever run on that fine-tune. You set it once, it's fine, and then it sits there for six months while you poke at it anyway, because it's on screen staring at you, implying a better number exists.

Put a per-model setting in a per-scene panel and you haven't given anyone control. You've given them a slot machine. Pull the lever, the output changes, nobody can tell whether it changed for the better, and everyone walks away convinced the model is inconsistent.

There are exactly two values where "I want tonight to feel different" is a coherent thing to want.

TemperatureHow flat the odds get

Raise it and unlikely words start looking plausible. This is the dial that turns competent prose into something stranger, for better and for worse.

Min-pThe floor under a word's odds

Cuts off options that are far less likely than the front runner, scaled to how confident the model is. It's what keeps a high temperature from tipping into nonsense, which is why the two move together or not at all.

Those two are what the Imagination slider moves. They travel as a pair because moving one without the other gets you either mush or a model that plays it safe all night. Show them as two independent numbers and you've invited exactly the wrong experiment.

So: one control, named for what you want instead of what it computes. Want a weirder evening? Drag it right.

The rest of the sampler stack isn't gone. It's somewhere better: with the software that actually owns the model.

LM Studio, Ollama, llama.cpp and KoboldCpp all let you set top-k, top-p, repetition penalty and the rest per model, and they remember it. Set it once against the model it belongs to, and every story you run through that model inherits it. That's the right scope for a per-model value, and it survives you switching front ends.

On a cloud service? The defaults were picked by people who've tested that model harder than you're going to. Leave them alone.

The switch above the slider is the whole philosophy in one control. Off means we send nothing and your model and runner decide. No overrides, no opinions from us.

That's a genuinely good place to sit, not a fallback. If you've already tuned a model you like, we should get out of the way.

You can have it. Run a local model, open your runner, and every parameter is sitting right there with a text box next to it. We're not hiding anything, and we're not stopping you.

We're just not putting them somewhere that implies you should be changing them mid-scene. For almost everyone, that trade has been a bad one.