NVIDIA Nemotron 3 Diarization on Baseten: 500+ real-time streams per RTX PRO 6000
baseten · x · 2026-09-24
NVIDIA Nemotron 3 Diarization is now available day 0 in Baseten's Model Library as an open, end-to-end streaming speaker diarization model replacing the usual segmentation-plus-clustering pipeline.
- A single RTX PRO 6000 sustains 500+ concurrent hour-long diarization streams (190 with transcription)
- Configurable algorithmic latency from 0.32s to 30.4s; as cheap as 1 cent per audio hour
- Labels up to 8 speakers in one pass, no clustering, unbounded audio length
- 100M-param transformer built on Streaming Sortformer; speaker activity also informs VAD, endpointing, and turn-taking for voice agents
Related event: NVIDIA Nemotron 3 Diarization Launches on Baseten with Day-0 Support(2 posts)→
More from Infra
- AI Gateway explained in 2 minutes: routing, auth, caching and observability — _jaydeepkarale · 2026-09-24
- Jev turns past LLM responses into a semantic cache to cut agent inference costs — MikkoH · 2026-09-24
- AI Compute Firm Fluidstack Opens 8 Roles Around 'Decision Engineering' for Gigawatt-Scale Infrastructure — MxMnr · 2026-09-24
- Multi-GPU monitor plugin released for DeepSeek Harness, seeking Windows testers — paulqq · 2026-09-24
- Yoniq Compute open-sources inference recipes for SGLang/vLLM, validated on up to 8x H200 — TheZachMueller · 2026-09-24
- Microsoft to invest over $10B across UAE, Saudi Arabia, Qatar and Kuwait by 2030 — Polymarket · 2026-09-24