Nvidia’s new diarization model forbids you from publishing what you measure

Nvidia posted Nemotron-3-Diarization-preview on 5 September under an evaluation licence that requires the company’s prior written consent before a user discloses benchmark results.

That clause is unusual and it is the one to read twice. Every other restriction on the page limits what you can build; this one limits what you can say about what you measured.

The rest is a familiar preview licence, and it is restrictive on its own terms. The model is “provided through an Early Access program solely for internal testing and evaluation”, so it cannot go to production. Use is confined to systems with Nvidia GPUs. Redistribution of the model or any derivative is prohibited. Outputs may not be used “to improve or develop another AI model”, which rules out distillation and rules out using it to label training data for something you own.

The model itself is a streaming sortformer that tags speakers as audio arrives rather than after a file is complete, which is the property a call-centre or meeting pipeline actually wants. Nvidia has not published a comparison against the diarization models teams run today, and under its own licence whoever runs that comparison may not publish it either.