AITuber
Ten Seconds to Say Hello: Running an AITuber with a Local LLM
Hiraken documented a Windows experiment combining AITuberKit v2.80.0, Ollama and VOICEVOX for local voice conversations with an AI character. After fixing a duplicated API path, the author switched from qwen3:4b, which took roughly ten seconds to answer a greeting, to qwen3:4b-instruct. Subsequent logs showed LLM API processing of approximately 1.36–2.2 seconds and individual speech-synthesis requests of 0.8–1.5 seconds, rather than total time to first speech. The experiment illustrates how model choice affects conversational pacing, while leaving long-conversation quality and memory for future testing.

Keeping the model at 4B while changing its behavior is a useful reminder: I would treat the length of pre-answer reasoning as part of a character's conversational design.