The first question most people ask about local models is “which one?” The better first question is “how much room do I have?” — because on a local machine, RAM is the real budget. A model that fits comfortably will always feel better than a bigger model that has to swap.
Size is a trade, not a rank
A small model you can run quickly can improvise for hours. A much larger model might write more sophisticated prose, but at the cost of pace. For interactive fiction, pace matters: the narrator has to answer while the scene still feels alive.
A slow narrator does not just test your patience. It drains the momentum a good story depends on.
Voice is a different axis
Different model families read differently. Some narrate in short, punchy beats. Others lean literary and slow. The right model is the one whose voice fits the world you are playing — a noir city and a fairy-tale forest rarely want the same narrator.
This is why FPV treats the narrator as a per-story setting rather than a fixed identity.
Start small, upgrade deliberately
Begin with a model your machine runs without effort. Play a few scenes. If the voice is wrong, try a different family at the same size before jumping up. Voice is usually cheaper to fix than size.
You do not need one model forever
Keep a fast daily model and a slower, richer model for the scenes that matter. The story — and its memory — stays the same either way. That is the point of separating the narrator from the world state.