NVIDIA PersonaPlex Tests Sales LoRA Limits

Hot take: this is less a “can the model do sales?” question than a data and deployment question; the model looks strong enough as a base, but the bottleneck is whether a tiny, narrow dataset can actually teach stable sales behavior without overfitting.

• –NVIDIA’s official release describes PersonaPlex as a 7B Moshi-based full-duplex model with 24kHz audio and explicit emphasis on conversational dynamics, interruption handling, and task adherence.
• –The official page says code and model weights were released, so the Reddit framing of “inference only” may be understating what builders can actually work with.
• –18 hours of positive-call data is plausibly enough for style or persona adaptation, but probably not enough to encode robust sales policy from scratch without regression risk. This is an inference.
• –If the goal is sub-250ms latency and natural barge-in, keeping the core speech model mostly intact and starting with prompting/conditioning is the lower-risk path. This is an inference.
• –8kHz telephony audio should be treated as a source-format constraint, not a feature upgrade; resampling may make it usable, but it won’t recover high-frequency detail that was never captured.
• –The strongest signal in the thread is that people want evidence from real production deployments of full-duplex voice models, not just benchmark claims.