What roleplayers actually run

There are two ways to rank models for story. One: have a judge (human or AI) read outputs and score them. Two: watch what thousands of roleplayers actually run, night after night, when every token costs them something.

OpenRouter's roleplay rankings are the second kind... usage share across millions of users, updated from real weekly traffic. Nobody performs for this leaderboard. It's just where the tokens went.

Three things jump out.

DeepSeek is half the hobby. Three DeepSeek models add up to nearly 48% of roleplay traffic. Not because a judge crowned them... because they write well enough and cost close to nothing, and at hundreds of turns a week, price IS a feature.

The winners are open-weight. DeepSeek, Gemma, gpt-oss, Hy3, MiMo... nearly the whole top of the board is models with published weights. Roleplayers, as a population, have quietly voted against the closed frontier for this use case.

A third of usage is "everything else." That long tail is the finetune scene... hundreds of community models, each with a small devoted audience. No other genre of AI usage looks like this. It's a hobby with TASTE.

Honesty section. Usage is shaped by price as much as quality... a free tier at #7 is not a coincidence. Popular doesn't mean best for YOUR scene: benchmark authors themselves note that RP-tuned models score poorly on formal writing evals precisely because they write casual, conversational prose... which is often exactly what a story needs. And a board like this measures cloud traffic, so the purely-local crowd (the koboldcpp basement, bless it) is invisible here entirely.

So read it as a shortlist generator, nothing more.

Put the two findings together... open weights winning, a massive finetune long tail... and the conclusion writes itself: the models roleplayers prefer are the kind you can RUN YOURSELF. The same DeepSeek-lineage and Gemma-lineage models sit on Hugging Face as GGUFs right now, quantized for real hardware, along with the entire long tail the cloud boards can't even see.

That's the mi·do·na bet in one sentence: this hobby was always headed home. Pick a model off the board (or off the long tail), load it in Ollama, LM Studio, koboldcpp, or llama.cpp, and the 2am session stops costing per token and stops leaving your machine.