NVIDIA Nemotron 3.5 Transcribe: what it is, how to run it

Nemotron 3.5 Transcribe is a free, open-weight speech-to-text model from NVIDIA. It takes recordings of up to an hour in 40 languages, labels who said what, translates, and answers questions about the audio, all from one checkpoint you can download or call through a hosted API for about 12 cents per hour of audio.

The launch posts

@NVIDIAAI opened its thread on October 28 with "Meet NVIDIA Nemotron 3.5 Transcribe" and described a promptable mixture-of-experts model with 30B total parameters, roughly 4B of them active. The demo shows it keeping track of speakers while several people talk over each other. In the second post, NVIDIA says it ranks first among open-weight models on Hugging Face's Open ASR English leaderboard, with a 5.33% average word error rate.

Twelve minutes later, @vllm_project announced "day-0 serving through vLLM + NeMo Speech," with weights in BF16, FP8 and NVFP4. For anyone who wants to test it, I'd weigh that post above the leaderboard claim, because it means the standard serving tools worked on launch day.

If you see two different sizes quoted, that's why: the Hugging Face model card lists 34B total and about 4.3B active because it counts the audio encoder, while NVIDIA's posts leave it out.

What it can do

The part I like most is the prompting. You tell it what you want in plain words. NVIDIA's launch blog gives examples like "Transcribe this recording and label each speaker" and "What decisions were made in this meeting?" You can also give it a list of terms to listen for, such as drug names, part numbers or staff names. In NVIDIA's medical case study, naming the language and alphabet in the prompt cut the word error rate from 30.8% to 23.8%, and that was before any fine-tuning.

How to try it

The quickest test needs no setup at all. Upload a clip to NVIDIA's demo Space on Hugging Face (nvidia/nemotron-transcribe) and see how it handles your own audio.

For a hosted API, Baseten, DeepInfra and Simplismart all serve it. DeepInfra's post uses the standard OpenAI client pointed at its endpoint, so a call takes a few lines of Python:

client = OpenAI(api_key=os.environ["DEEPINFRA_API_KEY"],
                base_url="https://api.deepinfra.com/v1/openai")
with open("support_call.mp3", "rb") as audio:
    t = client.audio.transcriptions.create(
        model="nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B",
        file=audio,
        prompt="Plumbing service call. Terms: Navien, tankless, PEX.")
print(t.text)

Ask for response_format="verbose_json" to get word timestamps, or srt/vtt for captions. DeepInfra says speaker attribution in its API is coming soon, and that it keeps no data after a request.

To run it on your own hardware, the model card's quick start targets a single H100 80 GB with vLLM. Its hardware table also lists cards down to an RTX 5090 and DGX Spark, depending on precision. It accepts WAV, FLAC, MP3 or OGG.

My read

This is the second strong transcription release this week. Fish Audio Realtime STT went after live phone calls at 36 cents an hour. Nemotron covers the other half of the job: recorded meetings, calls and voicemails, where you want the whole file done accurately with names attached. Mistral's Voxtral 3, out earlier in October, is the obvious open alternative to compare it against.

Price is what caught my eye. At 12 cents per hour of audio, a business that records 200 hours of calls a month would pay about $24 to transcribe them, and nobody needs a per-seat meeting-notes subscription. Since the weights are open, a clinic or law office that can't send audio to an outside vendor can run it on its own machine. Read the OpenMDW license before you build a product on it, and keep in mind that the benchmark numbers come from NVIDIA and the leaderboards it chose to cite.

A transcript doesn't do anything by itself, though. A file of who said what only pays off once it turns into a follow-up email, a CRM note or a flagged complaint. If you want to see where your calls and meetings could feed straight into that kind of follow-up, New Face Design's free process audit maps it out before you buy anything.