PersonaPlex is an NVIDIA voice model released openly on 15 January 2026. It works in full-duplex mode: it listens and speaks at the same time, so you can interrupt it, and it murmurs “uh-huh” as you talk and chooses when to chip in. The role is set with a text prompt, and the voice is cloned from an audio sample. The code is open under the MIT licence, and the weights come under an NVIDIA licence that allows commercial use. The limitations are serious: English only, a data-centre GPU, and a memory that lasts a few minutes of conversation.
What the model can do
PersonaPlex has 7 billion parameters, fine-tuned on top of Moshi, the open voice model from the French lab Kyutai. The key difference from an ordinary voice assistant is that there is no “speak, wait, get an answer” cycle. The model notices when it has been interrupted and adjusts, pauses, says “uh-huh” and “yeah”, and changes its intonation.
By NVIDIA’s own measurements on the Full-Duplex-Bench benchmark, the model handles 95% of interruptions correctly and responds to a change of turn in 0.17 seconds. These figures have not been independently verified.
The role is set with a single text prompt, and the voice with a short audio sample: 18 ready-made voices come with it, and you can plug in your own. NVIDIA shows examples with a teacher, a bank employee and an astronaut on Mars. The synthetic training dialogues, though, covered only two roles: assistant and customer support. Alongside them the model was trained on 7,303 real, unscripted conversations from the Fisher corpus. A teacher or an astronaut it works out on its own.
Licence
The code is released under the MIT licence. The weights come under the NVIDIA Open Model License: commercial use and derivative models are allowed, but this is not open source in the strict sense. To download the weights you have to accept the licence terms, your rights terminate if you bypass the built-in safety mechanisms, and you have to credit Kyutai for the Moshi base model.
Where it falls short
- English only. The model’s author explained that Moshi itself is not multilingual, and the natural feel of the conversation rests on a corpus of English dialogues in which each speaker was recorded separately. No other open corpora like that exist.
- Hardware. The officially supported GPUs are the A100 and H100 data-centre cards. In the MakeUseOf test on a gaming RTX 3090, the model took up about 19 GB of video memory; on 8 GB cards only quantised community builds will run.
- Short memory. The model’s context is 2,048 frames, which is less than three minutes of conversation: anything earlier drops out. In the MakeUseOf test this summer the model forgot its role after a few minutes, changed its name and started going round in circles. Since the November code update the role and voice hold for the whole conversation, but the memory of the conversation itself is a sliding window of about two and a half minutes: the model still does not remember what was said at the start.
What happened next
NVIDIA never released a new version of the model: the weights on Hugging Face have not changed since 2 March 2026. The code, however, is still being updated: on 6 November NVIDIA added support for endless conversations in which the role and voice hold to the end. The project is popular: by the end of September it had more than 10,000 stars on GitHub and almost 2.9 million downloads on Hugging Face. The community has put together lightweight versions for Mac and for running in a browser.
In June Kyutai fine-tuned PersonaPlex with reinforcement learning and released its own version — but under a non-commercial licence. The main competitor came from OpenAI: in July ChatGPT’s voice mode moved to GPT-Live, a full-duplex model, and in September it was opened to developers at $0.05 a minute. The difference is fundamental. GPT-Live runs in the cloud with search, memory and tools, but you can’t plug in a voice of your own. Standard access offers only ready-made voices; custom ones, according to the voice company Rime, are available to selected business customers by arrangement with OpenAI. PersonaPlex, on the other hand, runs on your own hardware and can speak in any voice you like, but it is only a conversation model, with no knowledge and no long-term memory.
The bottom line
PersonaPlex showed that a natural voice conversation with AI is possible in an open model, not just from the big cloud services. For an English-language product — a voice support bot, a training simulator, a game character — it is a workable foundation. For a Russian-language product it isn’t yet: the language isn’t there, and the authors give no timeline: there is neither a multilingual base model nor an open set of recorded conversations in the right format.
Another model released in the open: Qwen-Image-2.0.
