NVIDIA opens Nemotron 3.5 Transcribe, a 30B-A4B speech model for 40 languages, plus 463k-hour Granary v2
NVIDIA published open weights for Nemotron 3.5 Transcribe, a mixture-of-experts speech model with about 4B active parameters. One model handles transcription in 40 languages, translation to and from English for 31 of them, and labeling of up to eight speakers in hour-long recordings. Alongside it comes Granary v2, a speech corpus covering 48 languages.
Why it matters
One commercially usable open model replaces separate ASR, translation and diarization pipelines, and it tops the Open ASR leaderboard among open-weight systems.
How to try it
Nemotron 3.5 Transcribe is an open-weight (OpenMDW-1.1) speech model you can run yourself for transcription, translation and speaker-labelled transcripts of recorded audio. NVIDIA's reference setup is a single H100 80GB.
- Set up the NeMo Speech environment as described on the model card.
- Download the checkpoint from Hugging Face (nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B).
- Load it with NeMo's ASRModel.from_pretrained and call transcribe() on a WAV file.
- For serving, follow the model card's vllm serve command and send audio through the OpenAI-compatible chat API.
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained("nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B")
transcriptions = asr_model.transcribe(["file.wav"])