Mind Bender Simulator ships Qwen3.5 NPCs

This reads less like a model launch and more like a real-world stress test for what “good” local RP actually means: consistency, refusal style, and latency under pressure. The takeaway is that 4B can be usable with strong prompting, but the quality gap to 9B still shows up fast in long chats.

• –The prompt setup is intentionally minimal, so the model is being judged on raw persona adherence rather than RAG, tools, or scaffolding
• –A 20+ turn secrecy game is a harsher test than casual chat; models that “feel smart” often fail when they have to refuse in-character for many turns
• –The bottleneck is not just quality but first-token latency, which matters a lot for a game loop built around repeated dialogue
• –Community suggestions in the thread point toward small roleplay-tuned models like Gemma and Llama 3.1 8B variants, but compatibility with the current Unity/llama.cpp stack is a hard constraint
• –If the project keeps growing, the real win will be finding a smaller model that is good enough with prompt discipline, not just a bigger base model