Introduction

NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight speech model that folds transcription, speaker labels, translation and audio Q&A into one checkpoint. It transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13, and takes recordings up to one hour long.

It's open weights under the OpenMDW-1.1 license, which NVIDIA says allows commercial and non-commercial use. You can download it today from Hugging Face in BF16, FP8 and NVFP4, and run it locally with vLLM and NeMo Speech. It's built for recorded audio, not live streaming.

What Shipped

From the model card and NVIDIA's launch post on Hugging Face:

  • Size: a hybrid Mamba-Transformer mixture-of-experts model, 34B total parameters with about 4.3B active per token, audio encoder included. The repo is named 30B-A4B.
  • Context: 65,536 tokens, trained on recordings up to one hour, 16 kHz mono input.
  • Tasks: verbatim or clean transcription, transcription with your own vocabulary, speaker-attributed transcripts with up to eight speakers, word timestamps, speech translation, language ID and audio question answering.
  • Weights: BF16, FP8 and NVFP4, plus an optional DFlash2 draft model for faster decoding.
  • Hardware: the reference setup is one H100 80 GB. NVIDIA's compatibility table also lists the RTX 5090 in NVFP4, with GGML as one of its runtimes.
  • Dataset: NVIDIA also opened Granary v2, a speech dataset with 463,062 hours across 48 languages, along with the curation pipeline behind it.
  • Release channels: Hugging Face, ModelScope, NGC, build.nvidia.com and the NeMo Speech GitHub repo, all dated October 28.

You Pick the Transcript Style With a Prompt

This is the part that changes daily work. Instead of a fixed output, you tell the model what kind of transcript you want, in plain words. The model card's prompt guide gives these, and you can paste them as they are:

Audio transcription. Language: as spoken. Wording: omit fillers and accidental repetitions; keep the remaining spoken wording unchanged. Formatting: punctuation and capitalization. Output: transcript only.

Swap in a verbatim version when you need every "um" for a podcast edit:
Audio transcription. Language: as spoken. Wording: verbatim, including fillers and repetitions. Formatting: punctuation and capitalization. Output: transcript only.
For a translation, NVIDIA's format is `Please translate the given speech to <language>.` For questions about the recording, ask directly and turn reasoning on.

### Pro Tip

Keep the system prompt empty for transcription and translation, and use temperature 0 with reasoning off. Switch to temperature 0.6, top_p 0.95 and reasoning on only when you're asking questions about the audio. Those are the card's own recommended settings.

## The Numbers, With Caveats

These are NVIDIA's results, captured October 12 and published on the model card:

- **English ASR:** 4.71 mean WER across nine public Open ASR Leaderboard sets, against 4.94 for Qwen3-ASR-1.7B, the next best open-weight model. NVIDIA's launch post on X puts its live Open ASR Leaderboard average at 5.33% WER, ranked first among open-weight models.
- **Long-form ASR:** 10.98 mean WER, against 13.37 for Parakeet-TDT-0.6B-v3.
- **Multilingual ASR:** 4.62 mean WER across seven languages, against 6.49 for Qwen3-ASR-1.7B.
- **Multi-speaker meetings:** 15.80 mean cpWER on full AMI, CALLHOME, Mixer 6 and NOTSOFAR-1 sessions, against 16.83 for the best published pipeline.
- **Serving speed:** on one RTX PRO 6000 Blackwell GPU, NVIDIA's optimized NIM reaches up to 2,238 seconds of audio per second. The card says those optimizations aren't in the open release yet, and the open vLLM path is roughly two times slower.

Quality also varies by language. NVIDIA rates 13 languages, including English, Hindi, Korean and Spanish, as "Leading," and others such as Mandarin, Arabic and Tamil only as "Supported," which may need fine-tuning. Audio Q&A and reasoning were only evaluated in English.

## How to Run It Today

You'll need vLLM 0.32.0 and the NeMo Speech main branch. The card's setup, shortened:
git clone https://github.com/NVIDIA-NeMo/Speech.git nemo-speechcd nemo-speechuv venv --python 3.13 .venvUV_PROJECT_ENVIRONMENT=$PWD/.venv uv sync --extra asr --extra vllm --locked --no-install-projectuv pip install --python .venv/bin/python --no-deps -e .source .venv/bin/activatehf download nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B --local-dir ./NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
The card also installs PyAV and OpenCV, and then starts `vllm serve` with `VLLM_PLUGINS=nemo_speechlm` and a long list of flags. Copy that block straight from the [Quick Start](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B#quick-start) instead of retyping it. Once it's up, it speaks the OpenAI chat API on port 8000:
import base64from openai import OpenAIclient = OpenAI(base_url="http://127.0.0.1:8000/v1&quot;, api_key="EMPTY")audio = base64.b64encode(open("episode.wav", "rb").read()).decode()response = client.chat.completions.create( model="nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B", messages=[{"role": "user", "content": [ {"type": "text", "text": "Transcribe the following audio."}, {"type": "audio_url", "audio_url": {"url": f"data:audio/wav;base64,{audio}"}}, ]}], temperature=0, extra_body={"chat_template_kwargs": {"enable_thinking": False}, "skip_special_tokens": False},)print(response.choices[0].message.content)

Why It Matters for Audio Creators

Most podcast and video workflows still chain three tools: one to transcribe, one to split speakers, and one to translate. Nemotron 3.5 Transcribe does all three from one open checkpoint, on your own GPU, with recordings up to an hour. That means a full interview can come back as a clean, speaker-labelled transcript with word timestamps, ready for captions or a text-based edit.

The prompt-based styles are the quiet win. A clean version for show notes and a verbatim version for the edit now come from the same model, with no second pass. The big catch is hardware: this is a data-centre-class model, and a single consumer card needs the NVFP4 build.

Conclusion

Nemotron 3.5 Transcribe is NVIDIA's open-weight answer to the whole transcription stack: 40 languages, eight speakers, translation and audio Q&A in one 4B-active model you can run on one H100. If you transcribe long multi-speaker recordings, it's the open model to test this week.

—Anabel ♡