Nemotron 3 Diarization: an evaluation licence, not an open model
7 min read
On 24 August 2026, at 23:00:02 UTC, NVIDIA published nvidia/Nemotron-3-Diarization-preview on Hugging Face, a diarisation model — working out who spoke and when, inside a recording with more than one voice. We verified this ourselves, querying the public API and the individual files between 30 and 31 August. One detail in the metadata deserves a careful read: the pipeline_tag field declares voice-activity-detection — detecting whether someone is speaking, not who — while the tags that actually describe the model are different: speaker-diarization, streaming-sortformer, speaker-tagging. Anyone who reads the terms before downloading starts here: a generic label for a more specific job.
The repository is light: 399,652,947 bytes in total — just under 400 megabytes — in a single .nemo file, not in the many .safetensors shards of the language models we usually cover. library_name: nemo, gated: manual. At that moment — 22:44 UTC on 30 August, just past midnight in Italy — the API recorded 31 downloads and 22 likes.
One gate, two different things behind it
Without authenticating, the README.md file returns HTTP 200. Its content, though, is a single line: “For the detailed information, see the Overview subcard.” — and that file sits behind the gate. We queried that file, along with bias.md, privacy.md, safety.md, explainability.md and the weight file Nemotron-3-Diarization-preview.nemo: all of them return HTTP 401, with no account. The same holds for a file the card does not even mention, nemotron3_tts_8_open_voices_v18.mp4, perhaps a sample of the synthetic voices used in testing.
On 17 August we had checked a gated repository where only the weights sat behind the login, with the licence itself readable outside the gate. Here the gate is wider: inside it sit overview.md and the four other files just named — their names point to the same kind of ethics-review subcards, though none of the four appears in the README’s one public line. For a model that handles voice, and separates the people speaking, the privacy note is the one you would want to read before requesting access, not after you have already obtained it.
An evaluation licence, not an open model
license: other, license_name: nvidia-software-and-model-evaluation-license. The access form says so from its own heading, “Terms and Conditions”, and from the line introducing it: “This model is provided through an Early Access program solely for internal testing and evaluation.” The operational restriction follows, verbatim: “The model and its derivatives may be used only for internal test and evaluation, not in production, and only on systems with NVIDIA GPUs. You may not redistribute the model or derivatives, use their output or artifacts to improve or develop another AI model”.
The form also asks whoever would use it two things, through closed-choice menus. To the question “What is your intended use for this speaker diarization model?” the options are: meeting transcription, broadcast or media transcription, media commentary or content analysis, metadata generation, TTS training-data annotation, speech dataset annotation, research and benchmarking, plus a free-text field. To the question “What type of organization or affiliation do you represent?” one of the options, among others, is “Government or public-sector organization”. There is no need to guess who NVIDIA expects to be asking: the form asks directly.
There is nothing improper in this: an evaluation licence is a legitimate instrument, and NVIDIA states as much in its very name. The point is a different one: whoever downloads this repository expecting to try an open-weight model is instead accepting a vendor’s evaluation agreement, not an open licence. Worth knowing before clicking “Agree”, not after.
What you cannot say about your own trial
The public licence page — nvidia.com, HTTP 200 — closes its list of prohibited uses with this clause: “Disclose any evaluation or test results regarding the Licensed Materials or Derivatives without NVIDIA’s prior written consent.” For a company, that is a contractual constraint like many others. For a public body that has to justify a procurement choice, or for anyone under a transparency duty toward an oversight body, it is different: the outcome of your own evaluation — it works, it does not, how costly its errors are — cannot be put on the record in a public act without the written consent of whoever granted the licence.
Every derivative flows back to NVIDIA
The same page holds the clause that carries the most weight. Accepting the licence means granting NVIDIA this: “you grant NVIDIA and its affiliates a non-exclusive, perpetual, irrevocable, sublicensable, worldwide, royalty-free, fully paid-up and transferable license, under your intellectual property rights, to publicly perform, publicly display, reproduce, use, make, have made, sell, offer for sale, distribute (through multiple tiers of distribution), import, create derivative works of and otherwise commercialize and exploit at NVIDIA’s discretion any Derivatives created by or for you.” And right after: “You may, but are not required to, deliver Derivatives to NVIDIA.” — you do not even need to hand the derivative over: creating it is enough.
“Derivative Models” is not a term for the obvious case of retouched weights alone. The licence defines it broadly: it covers “any other model that is created or initialized by transfer of patterns of the weights, parameters, activations or output of such Licensed Materials to another model, […] including distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Licensed Materials for training the other model”. A smaller model, trained on synthetic data this one generates, would be a Derivative Model — and the grant to NVIDIA over it is already signed before training even begins.
Why the voice is not a side detail here
Telling speakers apart, when the goal is identifying a specific person rather than just separating conversational turns, touches data that the General Data Protection Regulation treats as biometric — the same filter we verified for a fingerprint reader at a workplace. If processing of this kind requires an impact assessment, that assessment has to run before you start, not after you have already been granted access: here it cannot, because the model’s privacy note sits behind the same HTTP 401 as the weights. We offer no legal classification of our own: whether your diarisation use falls into that category depends on the actual use, the recordings processed and who supplies them.
What we do not know
We did not obtain access and have not read overview.md, bias.md, privacy.md, safety.md, explainability.md: we report the HTTP status, not content we have not seen. We do not know whether gated: manual means a human review of every request or an automated process slower than plain auto: Hugging Face’s public documentation describes both values, not how long this particular repository takes. We have not run the model, and know nothing about its accuracy at telling voices apart, nor the content of the attached .mp4 file. We do not know how long the Early Access programme will last, nor whether the licence will one day shift toward a stable release.
How we solve this
Comply: before authorising a trial like this one inside an organisation, someone has to put five things in writing — who signs the licence acceptance on the entity’s behalf, which machines it runs on (here, NVIDIA GPUs only), what recordings the model is tested on — real ones, with people informed and the necessary consent, or synthetic ones generated for the occasion — what part of the outcome can be published or reported to an oversight body without the vendor’s written consent, and what happens to any smaller model born of this trial, given that the licence claims it as a derivative. This is the record a register of models in production should already anticipate, not improvise after the first download.
Decide: the same set-up holds models, data, contracts and decisions together in a single operating register, for large enterprises, defence, government and healthcare — with voice treated for what the processing makes of it — separating turns, or identifying people — inside the client’s own perimeter, not inside a third-party vendor’s evaluation programme.
See the service · Talk to an engineer
Has someone in your organisation already requested access to an Early Access model to try it on real recordings? Half an hour with one of our engineers, at no cost: we will go through the request form and the licence together before anyone signs it.