NVIDIA's Nemotron 3.5 Transcribe takes a prompt with the audio

What happened

NVIDIA published Nemotron 3.5 Transcribe on October 28, 2026, together with Granary v2, the speech dataset it was trained on. Along with the audio, the model takes a short written instruction, so one model can transcribe, translate, label who spoke, or pull the decisions out of a meeting.

It's a mixture-of-experts model with about 30 billion parameters, roughly 4 billion of them active for each token. The weights are on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B.

What it can do

• Transcribe 40 languages, translate 31 of them to or from English, and attribute speakers in 13.
• Take recordings up to an hour long with up to eight labeled speakers and word-level timestamps.
• Answer questions about the audio, in English only for now.

How the prompt changes the result

The instruction is plain language. NVIDIA's examples are as short as "Transcribe this recording and label each speaker" or "What decisions were made in this meeting?" The useful part is context: in a medical test, NVIDIA added a list of likely terms ("Possible terms: HbA1c, metformin.") and spelling rules, and the word error rate on Eka Care's medical set fell from 30.8% to 23.8% before any training. Fine-tuning on about 245 hours of medical and accented speech took it to 11.8%.

The numbers NVIDIA gives

Granary v2, the training set, holds 463,062 hours of speech in 48 languages, up from 25 languages in the first release. NVIDIA's post names no license for the model or the dataset; check the license on the model card before you build on it commercially.

What it means for you

1. Interviews, calls, lectures: list the names and terms the model will hear, as NVIDIA did in its medical test.
2. Your own machine: wait. vLLM and SGLang-Omni serve it today on data-center GPUs; the smaller builds for RTX and Jetson hardware and a C++ runtime are listed as coming soon.
3. No GPU at all: NVIDIA names Baseten, DeepInfra and Simplismart as hosts.
4. Live captions: pick another model. This one is built for recorded files, not streaming.

With a transcript in hand, you can summarize it into meeting notes or build a discovery brief from a sales call.

Try it

Transcribe this recording and label each speaker as Speaker 1, Speaker 2 and so on, with a timestamp at each change of speaker.
People you may hear: [names]
Terms you may hear: [product names, acronyms, jargon]
Keep these spellings exactly. Write numbers as digits. Mark anything you can't make out as [inaudible].

Below is a raw transcript of a meeting with speaker labels. Write the notes:
1) Decisions: what was decided and who agreed. Quote the line each decision comes from.
2) Action items: owner, task, deadline if one was said. Leave the deadline blank if none was said.
3) Open questions nobody answered.
Don't add anything that isn't in the transcript. If a speaker label looks wrong, say where.

Transcript:
[paste]

I'm going to transcribe recordings about the topic below. From the documents I paste, list every name, product, acronym and technical term a speech model is likely to get wrong. Group them as people, products, acronyms and jargon. For each acronym, add how people say it out loud. Keep the list under 60 items.

Topic: [topic]
Documents:
[paste an agenda, a slide deck's text or past notes]

Sources

• NVIDIA on Hugging Face, Nemotron 3.5 Transcribe and Granary v2, October 28, 2026
• Hugging Face, NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model page, checked October 28, 2026