<- all tokdocs

NVIDIA PersonaPlex-7B: Full-Duplex Voice AI, License and Setup

By , founder of Mavric and Apso

View on TikTok ->

PersonaPlex-7B is an NVIDIA speech-to-speech model that listens and talks at the same time, so a conversation does not have to wait for turns. A TikTok from @morenicksaraev calls it "just released" and "completely free, open source." The full-duplex claim holds, but NVIDIA released it on 15 January 2026, 278 days before the video, and only the code is open source; the weights sit behind a gated, NVIDIA-specific license.

What is PersonaPlex-7B?

The repository README calls PersonaPlex "a real-time, full-duplex speech-to-speech conversational model." The paper is titled "PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models," by Rajarshi Roy, Jonathan Raiman, Sang-gil Lee, Teodor-Dumitru Ene, Robert Kirby, Sungwon Kim, Jaehyeon Kim and Bryan Catanzaro.

The correct identifiers are nvidia/personaplex-7b-v1 on Hugging Face and NVIDIA/personaplex on GitHub. The video's auto-captions render the name as "Personaplex 7 Build," which will not turn up in a search.

Is PersonaPlex new?

No. NVIDIA's ADLR research page gives the release date as 15 January 2026. The GitHub repository was created on 2026-01-05, the first code commit is dated 2026-01-14, and the arXiv paper's first version appeared the same day. The video was posted on 2026-10-20.

The project has not changed much since. GitHub and Hugging Face both show the last update on 2026-03-02, and the repository has 22 commits. By the time of the video, the model had 101,806 Hugging Face downloads, 2,912 likes and 11,708 GitHub stars.

Is PersonaPlex free and open source?

Partly. The README says: "The present code is provided under the MIT license. The weights for the models are released under the NVIDIA Open Model license."

MIT is an open-source license. The NVIDIA Open Model License is not OSI-approved, though it does allow royalty-free commercial use and redistribution. It adds conditions:

  • Section 3.1 requires redistributions to include a "Notice" file reading "Licensed by NVIDIA Corporation under the NVIDIA Open Model License."
  • Section 2.1 ends your rights if you "bypass, disable, reduce the efficacy of, or circumvent any technical limitation, safety guardrail" without an equivalent replacement.
  • NVIDIA reserves the right to "update this Agreement to comply with legal and regulatory requirements at any time."

The weights are also gated. The Hugging Face API reports the repository as gated, and requests for the model files without logging in return HTTP 401. The README says to "Log in to your Huggingface account and accept the PersonaPlex model license," then set HF_TOKEN. There is no charge, but you need an account and must accept NVIDIA's license first.

How well does PersonaPlex handle conversation?

NVIDIA's ADLR page compares PersonaPlex with Moshi and Freeze Omni:

  • Conversation dynamics (higher is better): PersonaPlex 89.1%, Moshi 71.8%, Freeze Omni 75.5%.
  • Latency (lower is better): PersonaPlex 0.205 s, Moshi 0.266 s, Freeze Omni 0.257 s.
  • Task adherence (GPT-4o judge): PersonaPlex 4.29 on FullDuplexBench and 4.40 on ServiceDuplexBench, against Moshi at 4.34 and 0.77 and Freeze Omni at 3.62 and 1.75.

The gain is large on conversation dynamics and service tasks. On latency, PersonaPlex beats Moshi by about 23 percent.

How to set up PersonaPlex with a role and a voice

The video says to "give it a text prompt for the role you want it to play, and then give it an audio example of the voice you want it to use." Both work.

The role goes in --text-prompt. The README includes sample prompts such as "You work for CitySan Services which is a waste management and your name is Ayelen Lucero..."

The voice goes in --voice-prompt. Every README example uses one of 18 pre-packaged voice embeddings (NATF0-3, NATM0-3, VARF0-4 and VARM0-4) from voices.tgz on the gated Hugging Face repo, for example --voice-prompt "NATF2.pt". The code also accepts your own audio. In moshi/moshi/offline.py, any path that does not end in .pt goes to load_voice_prompt, which in moshi/moshi/models/lm.py loads an audio file and normalizes it to -24 LUFS. The module docstring describes loading "a voice prompt WAV."

Running it needs a GPU. The README tests on an NVIDIA A100 80GB, launches the server with python -m moshi.server --ssl "$SSL_DIR" after installing the Opus development library, and offers --cpu-offload "If your GPU has insufficient memory." The video shows no install step or hardware requirement.

Did NVIDIA build PersonaPlex from scratch?

No. Hugging Face lists the base model as kyutai/moshiko-pytorch-bf16, and the README says "PersonaPlex is based on the Moshi architecture and weights." It "finetunes Moshi and benefits from the generalization capabilities of the underlying Helium LLM." Kyutai built Moshi. NVIDIA's addition, per the paper, is "hybrid system prompts, combining role conditioning with text prompts and voice cloning with speech samples," trained on a large synthetic dataset.

The model is English only. ADLR notes the training data "only covers service and assistant scenarios," and the README says it was trained on synthetic conversations for assistant and customer-service roles, with casual conversation from the Fisher English Corpus.

Bottom line

PersonaPlex-7B is a real full-duplex voice model with strong published results against Moshi, which it is built on. It has been out since January 2026, the code is MIT, and the weights need a Hugging Face login and acceptance of NVIDIA's non-OSI license. Plan for a capable GPU and English-only, service-style conversations.

Sources

Published October 20, 2026. Based on a TikTok by @morenicksaraev. Full video notes (markdown)