Most people ask which model is best and expect a name. The honest answer is that the question is missing a word: best at what. A model that writes gorgeous slow-burn prose will lose the thread of a plot forty replies in, and the one that holds the plot will write like a competent report.
You do not have to decide once. The character, its personality and everything it remembers are separate from the model generating the words, so switching is a per-scene choice rather than a commitment.
What each kind is good at
Five are available, and they are genuinely different instruments rather than five sizes of the same thing.
- Standard — what every new character runs on. Quick, steady and good enough for most scenes. You never have to pick it; it is already there.
- GLM — a step up when a scene gets complicated. Tighter logic, and it holds a plan across a long exchange.
- Claude Opus — the richest prose and subtext. Slow-burn scenes, sharp restraint, characters that imply more than they say.
- Google Gemma — quick, warm and playful. Good for breezy banter and light back-and-forth.
- DeepSeek — holds long, twisting stories together without losing the thread.
How to tell you are on the wrong one
The symptoms are specific, and once you know them you stop guessing.
The replies are pretty but the plot has quietly stopped moving. Someone promised something three scenes ago and nobody has mentioned it since. That is a reasoning problem — move toward GLM or DeepSeek.
Everything is coherent and nothing lands. The character says exactly what it means, in order, with no subtext. That is a prose problem, and it is what Claude Opus is for.
The scene has gone stiff and formal when it should be light. A warmer, faster model fits better than a more capable one; this is where Gemma earns its place.
The character contradicts something established long ago. Check memory before blaming the model — if the fact was never written down, no model will invent it correctly.
Switching does not reset anything
The model is not part of the character. The persona, the memory, the scene and the history all stay exactly where they were; only the thing generating the next reply changes.
That makes switching cheap enough to use as a tool rather than a decision. A common pattern is a fast model for the back-and-forth of a scene and a switch to a richer one for the moment that matters — a confession, a confrontation, an ending.
One thing to expect: the voice will shift slightly. Two models given the same persona will not sound identical, and mid-paragraph the seam can be visible. Switching between scenes rather than inside one hides it almost completely.
Sampling matters more than people think
Model choice sets the ceiling; sampling parameters decide where in that range you land. Temperature controls how predictable a reply is — lower is steadier, higher is more surprising. The others (top_p, top_k, min_p, repetition penalty) constrain which candidate words the model considers at all.
A model that feels flat is often a temperature problem rather than a model problem. A character that repeats a favourite phrase every fourth reply is a repetition-penalty problem.
These are exposed per character and per package, and they export and import as JSON, so a set of values that works can be saved and moved rather than rebuilt from memory each time.
What the lineup is, and is not
Five models, chosen because they cover genuinely different ground rather than to make a number look impressive. The lineup changes as better models appear — it is a routing decision, not a promise about specific names.
There is no ranking here because the ranking depends entirely on the scene in front of you. A page that tells you one model is simply the best for roleplay has not tried a scene that needed a different one.