Every contestant in my reality TV game is a local on device LLM running through llama.cpp, and players can pick the model in realtime. So I pulled the same moment out of my nightly test runs: Piper, an ER night nurse, introducing herself at the season premiere. It's the same persona prompt on nine models, real unedited output, voiced afterwards with the game's own TTS.

Interesting findings:

  • SmolLM2 360M and Qwen2.5 0.5B get the facts right and stop there

  • Llama 3.2 1B (the current default) starts using her backstory

  • Ministral 3 3B had the best character beat ("knitting fast when people get loud")

  • Qwen3 4B turned the same backstory into a flirt ("my type?")

Is there any models you'd like to see next?

The game is Villa de Drama, if you want to follow along: https://store.steampowered.com/app/5462320/Villa_de_Drama/