01
What is Nemotron 3 Diarization?
NVIDIA Nemotron 3 Diarization is an open-weight speaker-diarization model of roughly 100 million parameters. Diarization answers who spoke when by assigning anonymous speaker channels and timestamps; it does not transcribe the spoken words or identify real people by name.
The final checkpoint supports up to eight speaker channels and preserves overlapping speech by allowing multiple channels to be active in the same frame. That makes it relevant to meetings, calls, interviews, podcasts and voice-agent memory.
02
Final release, license and availability
NVIDIA's final model card lists a September 23, 2026 release and publishes nvidia/Nemotron-3-Diarization on Hugging Face under OpenMDW 1.1 for commercial or non-commercial use.
This final license differs materially from the earlier preview terms. At verification time, no managed Hugging Face Inference Provider was shown on the model page, so the documented path is local or self-managed inference rather than an assumed hosted API.
03
One checkpoint for offline and streaming diarization
NVIDIA documents recommended input-buffer configurations of 30.4 seconds for offline-style processing and 1.04, 0.64 and 0.32 seconds for progressively lower-latency streaming. The architecture can operate below that, but 0.32 seconds is the lowest recommended profile.
These numbers describe buffered audio, not complete application latency. Compute, transport, ASR and downstream processing add delay. Streaming continuity is preserved through speaker-cache and recent-frame context carried across chunks.
04
Benchmark results and what they do not prove
NVIDIA and Hugging Face report an initial Voice Arena Diarization-Bench result of 14.72% DER across 139 English conversations totaling about 22 hours, compared with 19.3% for the next-ranked system in that initial evaluation.
The result is explicitly preliminary, while NVIDIA's own tables report about a 41% unweighted mean relative DER reduction versus its earlier four-speaker Streaming Sortformer at a 1.04-second input buffer. These are reported benchmark results, not guarantees for every acoustic environment.
05
Input, hardware and practical limits
The reference workflow expects 16 kHz mono audio and can return speaker activity at fine time resolution. The eight-speaker maximum is a real design boundary, and unusually long recordings, noise, reverberation, far-field capture or domain shift can reduce quality.
NVIDIA does not publish one universal minimum RAM or VRAM requirement in the reviewed final model card. Parameter count alone is not enough to promise a device requirement; runtime, precision, batch size, streaming state and any accompanying ASR model all matter.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.