LM Studio, Ollama falter on roleplay

This looks less like "uncensored models are broken" and more like a stack mismatch: roleplay needs model quality, memory, and formatting discipline, not just permissive weights. 8B instruct models often lose character state over long chats, prompt structure and stop rules matter more than raw length, conservative context or quant settings can make generations incoherent, RP-first frontends like SillyTavern handle lorebooks and memory better than plain text-file hacks, and mid-size RP-tuned models usually beat generic 8B instruct variants for consistency.