<- all tokdocs
2026.10.20@morenicksaraev

PersonaPlex-7B really is full duplex, but NVIDIA shipped it on 15 January 2026, 278 days before this video said "just released," and the weights are gated behind a license the OSI does not recognize

AI
Watch on TikTok

View on TikTok ->

A 22.100998 s MP4 (container mov,mp4,m4a,3gp,3g2,mj2, HEVC Main video at 1080x1920, 30 fps, 663 frames, 298,739 bps; HE-AACv2 audio at 44,100 Hz stereo, 64,549 bps; 370,519 bps overall, 1,023,607 bytes on disk), posted by uploader morenicksaraev (channel morenicksaraev, uploader id 7695359986843239456) at epoch 1792488643, which is 2026-10-20T09:30:43Z, credited to "original sound" by morenicksaraev, with a 415-character description. At the capture time recorded on metadata.json (27 October 2026, 00:05 local) the post showed 90,500 views, 4,132 likes, 27 comments, 1,280 reposts and 4,850 saves. I read all 11 extracted frames (frame_001 through frame_011, roughly 0 s, 2 s, 4 s, 6 s, 8 s, 10 s, 12 s, 14 s, 16 s, 18 s and 20 s) and the full 96-word, 5-segment transcript. The frames run a continuous talking-head cut against custom motion graphics with no gap in the caption sequence, so I did not re-extract denser. Frame 001 carries the NVIDIA wordmark over a green waveform card and the burned-in caption "just released"; frame 002 shows a dark card with a microphone row labelled "Listens" and a speaker row labelled "Talks" under the caption "AND TALKS"; frame 003 is the presenter alone against a grey seamless backdrop with the caption "to it feels"; frame 004 shows a chip reading "nvidia / personaplex-7b-v1" above the word "PersonaPlex" with the caption "PERSONAPLEX 7B"; frame 005 lists three rows, "$0 Completely free" with a green check, "Open source / github.com/NVIDIA/personaplex", and "From NVIDIA", under the caption "COMPLETELY FREE"; frame 006 is the presenter with "of having to"; frame 007 shows a card headed "Taking turns" with alternating "You" and "AI" blocks under "the conversation"; frame 008 shows a "Text prompt" field with a single typed "Y" and a cursor, captioned "A TEXT PROMPT"; frames 009 and 010 show an "Audio example" player labelled voice_sample.wav feeding a green waveform labelled "PersonaPlex talks in this voice"; frame 011 is the presenter pointing down under "just check the". No frame shows the real Hugging Face page, the real GitHub repository, a terminal, or the model actually running, and no sample of PersonaPlex output audio plays at any point. The spoken argument is that NVIDIA has released a free, open-source voice model that listens and speaks at the same time so conversation flows without turn-taking, that you steer it with a text prompt for the role and an audio clip for the voice, and that a pinned comment holds the link. The last transcript segment ends at 21.94 s against an audio stream of 22.100998 s and an extracted audio.mp3 of 22.177959 s, so every segment closes inside the real audio and nothing in this transcript is a past-the-end hallucination.

The product name in the transcript is a Whisper mangle: "Personaplex 7 Build" is PersonaPlex-7B

At 6.20 s the transcript reads "It's called Personaplex 7 Build, and it's completely free, open source, and from Nvidia." The phrase "7 Build" is a Whisper transcription artifact for "7B." Two independent on-artifact checks settle it. The video's own description, which is typed rather than transcribed, reads "It's called PersonaPlex-7B." The burned-in graphic in frame 004 at roughly 6 s shows the repository identifier nvidia / personaplex-7b-v1 in a chip above the styled wordmark "PersonaPlex," with the caption "PERSONAPLEX 7B" underneath.

The Hugging Face API confirms the canonical identifier. A request to https://huggingface.co/api/models/nvidia/personaplex-7b-v1 returns "modelId":"nvidia/personaplex-7b-v1" with "pipeline_tag":"audio-to-audio" and "library_name":"moshi". The published paper title is "PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models." Anyone searching for "Personaplex 7 Build" from the transcript alone will find nothing. The correct strings are nvidia/personaplex-7b-v1 on Hugging Face and NVIDIA/personaplex on GitHub.

"Just released" is wrong by 278 days, and the repository has been quiet for 232 of them

At 0.00 s the video opens "Nvidia just released a free voice AI that listens and talks at the same time," and frame 001 burns "just released" across the screen. The video was posted at epoch 1792488643, which is 2026-10-20T09:30:43Z. The dated primary sources put the release nine months earlier.

The GitHub API reports that NVIDIA/personaplex was created at 2026-01-05T19:10:35Z. The oldest commit in the repository history is dated 2026-01-14T22:48:51Z with the message "Adding PersonaPlex inference code." The arXiv API reports the paper's first version published at 2026-01-14T07:47:46Z. NVIDIA's own ADLR research page gives the release date as 15 January 2026. Measured from that release date, the video is 278 days late. Measured from the first code commit, it is 279 days late.

The repository has not moved since. The GitHub API gives pushed_at as 2026-03-02T16:07:10Z, and the Hugging Face API gives lastModified as 2026-03-02T18:03:29.000Z. That is 232 days before the video was posted. The whole repository is 22 commits and 143 files. This is a settled January release with a quiet March touch-up, not news.

The scale of the existing adoption makes the framing harder to defend. By the capture date the Hugging Face repository reported 101,806 downloads and 2,912 likes, and the GitHub repository carried 11,708 stars. A model with six figures of downloads behind it was not released the week this video went up.

"Completely free, open source" is half right: the code is MIT, the weights are not open source and are gated

At 6.20 s the video says the model is "completely free, open source, and from Nvidia," and frame 005 at roughly 8 s presents this as three checked rows, headed by "$0 Completely free" with a green tick. The repository's own LICENSE section splits the claim in two: "The present code is provided under the MIT license. The weights for the models are released under the NVIDIA Open Model license." The GitHub API confirms the repository license as MIT. The Hugging Face card metadata confirms the weights as license: other, license_name: nvidia-open-model-license.

MIT is open source. The NVIDIA Open Model License is not an OSI-approved open source license. It permits commercial use royalty-free and permits redistribution, which is more permissive than many model licenses, but it attaches conditions that OSI-approved licenses do not. Section 3.1 requires that redistributions "include the following attribution notice within a 'Notice' text file with such copies: 'Licensed by NVIDIA Corporation under the NVIDIA Open Model License'." Section 2.1 terminates your rights if you "bypass, disable, reduce the efficacy of, or circumvent any technical limitation, safety guardrail or associated safety guardrail hyperparameter" without substituting an equivalent safeguard, and it reserves NVIDIA's right to "update this Agreement to comply with legal and regulatory requirements at any time." A license the licensor can unilaterally revise is not what "open source" means.

The weights are also gated, which the "$0" card with its green tick does not convey. The Hugging Face API returns "gated":"auto" for this repository. Unauthenticated requests for the files return HTTP 401: https://huggingface.co/nvidia/personaplex-7b-v1/raw/main/README.md returns 401 with the body "Access to model nvidia/personaplex-7b-v1 is restricted. You must have access to it and be authenticated to access it. Please log in," and a HEAD request for model.safetensors returns 401 as well. The project README states the requirement plainly: "Log in to your Huggingface account and accept the PersonaPlex model license," then export HF_TOKEN=<YOUR_HUGGINGFACE_TOKEN>. There is no charge, so "$0" holds. "Completely free" with no caveat does not: you need a Hugging Face account and you must accept NVIDIA's license before a single byte of the weights moves.

The full-duplex claim checks out, and the benchmark numbers support it

The central claim survives. At 0.00 s the video says the model "listens and talks at the same time," and at 10.68 s it says "instead of having to awkwardly take turns, the conversation just flows in real time." Frame 002 at roughly 2 s renders this as parallel "Listens" and "Talks" waveform rows, and frame 007 at roughly 12 s shows a "Taking turns" card with interleaved "You" and "AI" blocks as the thing being replaced.

The paper's title is "PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models." The repository README describes it as "a real-time, full-duplex speech-to-speech conversational model." NVIDIA's ADLR research page publishes the measured comparison. On conversation dynamics, where higher is better, PersonaPlex averages 89.1 percent against Moshi at 71.8 percent and Freeze Omni at 75.5 percent. On latency, where lower is better, PersonaPlex averages 0.205 seconds against Moshi at 0.266 and Freeze Omni at 0.257. On task adherence scored by a GPT-4o judge, PersonaPlex scores 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 video quotes none of these figures, so there is nothing to correct here. The numbers are worth recording because they show the margin is real on conversational dynamics and task adherence, and narrower on latency, where PersonaPlex's 0.205 seconds beats Moshi's 0.266 seconds by about 23 percent rather than by an order of magnitude.

The text-prompt and audio-sample workflow is real, and the code accepts an arbitrary WAV

At 14.62 s the video says "Just 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." Frame 008 shows a "Text prompt" entry field and frames 009 and 010 show an "Audio example" player labelled voice_sample.wav producing a waveform captioned "PersonaPlex talks in this voice." Both halves hold up, though the second needs a note.

The role half is documented directly. The README's offline example passes --text-prompt "$(cat assets/test/prompt_service.txt)" and publishes sample role prompts verbatim, including "You work for CitySan Services which is a waste management and your name is Ayelen Lucero. Information: Verify customer name Omar Torres. Current schedule: every other week. Upcoming pickup: April 12th. Compost bin service available for $8/month add-on."

The voice half is subtler than the README alone suggests. Every documented example passes a pre-packaged embedding, as in --voice-prompt "NATF2.pt", drawn from a fixed labelled set the README lists: NATF0 through NATF3, NATM0 through NATM3, VARF0 through VARF4 and VARM0 through VARM4, eighteen voices in total, shipped as voices.tgz on the gated Hugging Face repository. Read only that, and the video's voice_sample.wav card looks like an invention. The source code says otherwise. In moshi/moshi/offline.py the branch at line 236 reads if voice_prompt_path.endswith('.pt'): and calls load_voice_prompt_embeddings(voice_prompt_path), falling through to load_voice_prompt(voice_prompt_path) for anything else. In moshi/moshi/models/lm.py line 960, load_voice_prompt calls load_audio(voice_prompt, self._sample_rate) and normalizes the result to -24 LUFS. The module docstring at line 33 describes the "Prompt phase: load system text tokens and a voice prompt WAV (agent side)." Passing your own WAV is supported in code. It is just not the path the README demonstrates.

NVIDIA released it, but the architecture and base weights are Kyutai's Moshi

Frame 005 lists "From NVIDIA" as a standalone selling point beside the NVIDIA logo, and the video repeats "and from Nvidia" at 6.20 s. NVIDIA did release PersonaPlex, and the eight authors on the paper are Rajarshi Roy, Jonathan Raiman, Sang-gil Lee, Teodor-Dumitru Ene, Robert Kirby, Sungwon Kim, Jaehyeon Kim and Bryan Catanzaro. The causal order underneath that deserves recording, because the video's framing implies NVIDIA built a full-duplex speech model from nothing.

The Hugging Face metadata names the parent explicitly: base_model: kyutai/moshiko-pytorch-bf16 and base_model:finetune:kyutai/moshiko-pytorch-bf16, with library_name: moshi. The README states it outright: "PersonaPlex is based on the Moshi architecture and weights," linking Moshi's own paper at arXiv 2410.00037. The repository's generalization section adds that "Personaplex finetunes Moshi and benefits from the generalization capabilities of the underlying Helium LLM." Kyutai shipped Moshi first and PersonaPlex is a finetune of it. NVIDIA's contribution, per the paper abstract, is the persona layer: "hybrid system prompts, combining role conditioning with text prompts and voice cloning with speech samples," trained on "a large-scale synthetic dataset of paired prompts and user-agent conversations."

Two practical limits round this out, neither mentioned in the video. The Hugging Face tags list en as the only language. The ADLR research page notes that the training data "only covers service and assistant scenarios," and the README confirms the model "is trained on synthetic conversations for a fixed assistant role and varying customer service roles," with casual conversation coverage from the Fisher English Corpus. Running it needs a GPU. The README prescribes installing the Opus development library and launching python -m moshi.server --ssl "$SSL_DIR", and offers a --cpu-offload flag "If your GPU has insufficient memory." The video shows no terminal, no install step and no hardware requirement at any point in its 22 seconds.

Key Takeaways

  • Correction: The video says at 0.00 s that NVIDIA "just released" the model. NVIDIA's ADLR research page dates the release to 15 January 2026 and the oldest GitHub commit is 2026-01-14T22:48:51Z. The video was posted 2026-10-20T09:30:43Z, 278 days after release.
  • Correction: The video calls it "completely free, open source" at 6.20 s and frame 005 ticks "$0 Completely free." The README splits it: "The present code is provided under the MIT license. The weights for the models are released under the NVIDIA Open Model license." The NVIDIA Open Model License is not OSI-approved, it requires an attribution notice file, it terminates on circumventing safety guardrails, and NVIDIA may update it unilaterally.
  • Correction: The weights are gated. The Hugging Face API returns "gated":"auto", and unauthenticated requests for README.md and model.safetensors both return HTTP 401. The README instructs you to log in and accept the license before setting HF_TOKEN. No frame or spoken line mentions this.
  • Correction: The transcript's "Personaplex 7 Build" at 6.20 s is a Whisper artifact. The correct name is PersonaPlex-7B, repository nvidia/personaplex-7b-v1, confirmed by the burned-in chip in frame 004 and by the Hugging Face API.
  • Verified: The full-duplex claim holds. The paper title is "PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models" and the README calls it "a real-time, full-duplex speech-to-speech conversational model."
  • Verified: The text-prompt-plus-audio-sample workflow is real. moshi/moshi/offline.py exposes --text-prompt and --voice-prompt, and load_voice_prompt in moshi/moshi/models/lm.py line 960 loads an arbitrary audio file rather than only a packaged .pt embedding.
  • Clarification: Every documented voice example uses one of eighteen pre-packaged embeddings (NATF0-3, NATM0-3, VARF0-4, VARM0-4) shipped in voices.tgz. Supplying your own WAV works but is not the path the README demonstrates.
  • Clarification: Frame 005's "From NVIDIA" row understates the lineage. Hugging Face metadata lists base_model: kyutai/moshiko-pytorch-bf16 and the README states PersonaPlex "is based on the Moshi architecture and weights." Kyutai's Moshi came first; NVIDIA added the persona and voice-control layer.
  • Context: ADLR's published benchmarks give PersonaPlex 89.1 percent average on conversation dynamics against Moshi's 71.8 percent, 0.205 s average latency against Moshi's 0.266 s, and GPT-4o judge task-adherence scores of 4.29 on FullDuplexBench and 4.40 on ServiceDuplexBench. The video quotes no figures.
  • Context: English only per the Hugging Face en tag, trained for assistant and customer-service scenarios, and GPU-bound. The README tests on NVIDIA A100 80GB and provides --cpu-offload for smaller GPUs. The video shows no install, no terminal and no hardware note.
  • Context: The project has not been touched since 2026-03-02, 232 days before the video, and already carried 101,806 Hugging Face downloads, 2,912 likes and 11,708 GitHub stars at capture time.
  • Unverified: The video directs viewers to "the pinned comment below" at 19.46 s for a link. The pinned comment is not part of the downloaded artifact, so its destination cannot be checked from these files.

Resources

Published October 20, 2026. Writeup generated from a favorited TikTok.