Nvidia has open-sourced its PersonaPlex voice AI model under permissive licences, allowing developers to self-host commercially deployable real-time conversational AI without royalties or vendor lock-in.
Nvidia has open-sourced PersonaPlex, a 7-billion-parameter full-duplex voice AI model that can listen and speak simultaneously, making it freely available for commercial deployment under permissive licences. The code is released under the MIT licence, while the model weights are distributed under the Nvidia Open Model License, allowing commercial use without royalties and with no claim by Nvidia over users’ outputs.
Built on Kyutai’s open-source Moshi conversational model, PersonaPlex enables developers to self-host voice AI instead of relying on proprietary APIs. The release lowers the barrier for startups and enterprises building customer service, reception and other conversational AI applications.
According to Nvidia, PersonaPlex achieves a speaker-switch latency of 0.07 seconds, compared with 1.3 seconds for Google’s Gemini Live in the cited tests. The full-duplex architecture allows the model to listen and respond simultaneously, learning conversational behaviours such as pausing, interrupting appropriately and producing listener acknowledgements.
The model takes a short audio sample to recreate a voice and a text prompt defining the conversational role. Benchmarks cited by The Decoder report higher naturalness, 99.2% speaker-change success and improved voice similarity compared with Qwen 2.5 Omni and Moshi.
PersonaPlex was trained using 7,303 real conversations and more than 144,000 synthetic dialogues generated with Qwen3-32B, GPT-OSS-120B and Resemble AI’s Chatterbox. The English-only model is available on GitHub and Hugging Face, where its model card reports about 225,000 downloads over the past month. Nvidia plans further post-training alignment and tool integration as part of its broader open-model strategy following Nemotron 3 Nano.
