NVIDIA just dropped PersonaPlex-7B, a 7-billion-parameter AI model that fundamentally changes how machines talk to humans. Released on January 15, 2026, this full-duplex conversational model eliminates the awkward choice between natural-sounding AI and customization that’s plagued voice assistants for years.
Traditional voice assistants work like walkie-talkies. You speak, they process through a clunky ASR-to-LLM-to-TTS pipeline, then they respond. There’s delay, no real interruptions, and conversations feel robotic.
Full-duplex models made things feel natural by letting AI listen and speak simultaneously, but locked developers into fixed voices and personalities. PersonaPlex changes this completely.
Real conversations with any persona you want
PersonaPlex handles interruptions, backchannels like “uh-huh” and “oh,” and authentic conversational rhythm while letting you define any role through text prompts. Want a wise teacher? A customer service agent handling a bank fraud case? An astronaut dealing with a reactor meltdown on a Mars mission? PersonaPlex delivers all of these scenarios naturally.
The model supports diverse voice options pre-packaged as Natural (NATF and NATM families) and Variety (VARF and VARM families), with both female and male options.
Before conversations begin, PersonaPlex is conditioned on two prompts: a voice prompt establishing vocal characteristics, and a text prompt specifying persona attributes like role, background, and scenario context. These combined inputs create a consistent conversational identity that stays stable throughout interactions.
Built on the Moshi architecture with around seven billion parameters, PersonaPlex uses a base model called Helium for language understanding, enabling reasoning and generalization beyond its training data.
The model operates on continuous audio encoded with a neural codec, predicting both text and audio tokens autoregressively. This single unified model replaces separate listening, thinking, and speaking components, preserving conversational flow without generating text first.
Open source and ready for real use
NVIDIA released PersonaPlex-7B-v1 under the MIT license for code and the NVIDIA Open Model License for weights, making it ready for commercial use.
The model is available on Hugging Face and GitHub with full documentation. Running it requires a GPU with at least 24 GB of VRAM like an RTX 3090/4090, and it operates at 24kHz sample rate for audio input and output.
The v1 release is English-only for now, though NVIDIA indicated support for other languages like Spanish is on the roadmap. Training used a combination of synthetic and real conversations, teaching the model not just sentence generation but actual conversational behavior, when to pause, interrupt, or backchannel.
Demonstrations show PersonaPlex handling general knowledge questions with natural interruptions, banking scenarios with empathy and accent control, technical spaceship emergencies with appropriate urgency, and casual conversations with contextual backchannels.
The model generalizes well to new and unexpected situations, maintaining its assigned persona even when topics change or situations become stressful.
Have you tried any full-duplex voice AI models yet, or are you excited to test PersonaPlex? Drop your thoughts below!