NVIDIA Nemotron 3 Diarization is here, and it's powering FluidMeet

Today NVIDIA released Nemotron 3 Diarization, its new open-weight model for working out who spoke when. We're one of the first to ship it, with day-zero support in FluidMeet, the new meeting memory inside FluidVoice.

FluidMeet records your meeting, transcribes it, and tells you who said what. Everything happens on your Mac. No bots. No cloud. No subscription.

Meet Nemotron 3 Diarization

Nemotron 3 Diarization is built for real conversations: meetings, calls, and podcasts, with people talking over each other.

• Up to eight speakers, labeled in the order they first speak.
• Streaming or offline from a single model, with latency that can go as low as 80 ms.
• No limit on recording length. A one-hour all-hands works the same as a five-minute stand-up.
• Compact at 100M parameters, small enough to run locally.

It's also far more accurate. On NVIDIA's meeting benchmarks, it makes up to two-thirds fewer speaker errors than NVIDIA's previous streaming diarization model.

Why speakers matter

Who owns the action item? Who said no? Who raised the escalation? A transcript without speakers can't answer any of that.

Real meetings are also messy. People interrupt, agree mid-sentence, and talk at the same time. Nemotron 3 handles this: when two people speak at once, both get credit for what they said.

How FluidMeet uses it

FluidMeet uses two models, each with one job.

NVIDIA Parakeet writes down what was said, word by word, with a timestamp on every word.

NVIDIA Nemotron 3 Diarization tracks who was talking at every moment.

FluidMeet then matches the two. Each word is tagged with whoever was speaking when it was said. Instead of a wall of text, you get a transcript that reads like a script:

Speaker 1 · 12:04 — Can we ship this by Friday?

Speaker 2 · 12:07 — Not without cutting the export feature.

Your choice of speech engine

Parakeet v3 is the default. If you prefer another engine, you can switch to other local models from NVIDIA or to OpenAI's Whisper. Nemotron 3 Diarization handles the speakers either way.

Private. Local. Yours.

Meetings are where the sensitive conversations happen: unreleased roadmaps, customer calls, hiring decisions, and financials. None of that should have to leave your laptop.

In FluidMeet, the recording, the transcript, and the speaker timeline are all processed and stored on your Mac. There's no server, no bot joining your calls, and no monthly fee. That makes it a natural fit for teams in regulated industries.

Nemotron 3 Diarization runs on-device through Core ML, so FluidMeet works on any Apple Silicon Mac (M1 or later) running macOS 15 or later, including the 2020 M1 MacBook Air.

What's next

Coming next month: meeting summaries and the ability to bring your meeting memory into Claude and Codex, whenever you choose to.

Try it today

FluidMeet, powered by NVIDIA Nemotron 3 Diarization, is available now in the FluidVoice beta for Mac.

Know what was said. Know who said it. Keep it on your Mac.