Nvidia Releases PersonaPlex-7B as an Open-Weights Model That Can Listen and Speak at the Same Time
On January 15, 2026, Nvidia released the weights and code for PersonaPlex-7B, an open conversational model built for full-duplex speech — it listens and responds in the same moment, instead of waiting for the other party to finish. The release moves full-duplex voice AI from API-based demos to something developers can actually download and run themselves.
What Was Launched
According to Startup Fortune, Nvidia published PersonaPlex-7B-v1 on January 15, 2026, with weights on Hugging Face and code on GitHub the same day. The model has 7 billion parameters. The code is licensed under MIT, while the weights fall under Nvidia's own NVIDIA Open Model License — a distinction that means the code can be used almost freely, while the terms governing the weights themselves may include restrictions not yet documented in the available material. The source describes the model as running on a single GPU rather than a server rack, but does not state which GPU models, quantization levels, or sampling rates are required.
Why Full-Duplex Is Different
Traditional voice AI consists, as Startup Fortune describes it, of three separate systems chained together: automatic speech recognition turns your voice into text, a language model writes a response, and a text-to-speech engine reads the response aloud. This chain introduces delays at every step, and it forces turn-taking where only one party speaks at a time — the system must wait until you are finished before it begins to think.
PersonaPlex breaks with this. According to the source, the model uses a dual-stream design: one stream continuously processes the audio coming in from the user, the other continuously generates the agent's speech and text, and both streams share the same underlying model state. This means the model can perceive that you are interrupting while it is speaking — and react as it happens — which is what allows the conversation to resemble a real conversation rather than an alternating playback of questions and answers.
The Moshi Legacy
The architecture is not Nvidia's invention. According to Startup Fortune, PersonaPlex builds on the dual-stream Transformer architecture from Moshi, the open conversational model from the French lab Kyutai, with Nvidia fine-tuning and extending it into PersonaPlex. That means the fundamental concept has been available in open form for some time; what is new in this release is the combination of Nvidia's fine-tuning, persona control, and licensing.
How the Voice and the Person Are Controlled
Before a conversation starts, PersonaPlex takes in two prompts, according to the source: a voice prompt built from audio tokens that sets the vocal characteristics and speaking style, and a text prompt that defines the persona itself — role, background, and the scenario it operates in. In other words, the developer can control both how the agent sounds and who it is, without training a new model.
The Numbers — With Caveats
Nvidia reports three key figures for PersonaPlex-7B:
- 90.8 percent turn-taking rate — how often the model handles conversational turn-taking the way humans do.
- 240 milliseconds interruption response — how quickly the model reacts when the user cuts in.
- Around 170 milliseconds time-to-first-token — how quickly the model begins to respond.
These figures are Nvidia's own, relayed by Startup Fortune, and have not been independently verified. It has also not been confirmed that the release includes a technical report or reproducible benchmark code that would let developers check the figures locally, despite the source suggesting something to that effect. The methodology behind the evaluations is unknown in the available material.
What Remains Unclear
Several practical questions remain. There is no confirmed primary documentation from Nvidia — no model card, technical report, Hugging Face page, or GitHub repository — in the material this story is based on; everything stems from a single secondary source. Which languages the model supports, what training data it was built on, and what restrictions the NVIDIA Open Model License in practice places on commercial use are not detailed. Nor are the hardware requirements beyond "a single GPU" specified.
For developers considering open voice AI, the main point is nevertheless clear: full-duplex conversation technology, previously something encountered mainly through APIs and demos, now exists as a downloadable open-weights model with open code — provided the documented details hold up once primary sources become available.