Nvidia releases Nemotron 3.5 Transcribe for recordings in 40 languages
Nvidia released downloadable weights for Nemotron 3.5 Transcribe on 28 October 2026. The model handles up to one hour of recorded audio in 40 languages, including French, according to its official model card.
It supports transcription, speaker labels, word timestamps and selected translation or audio-question-answering tasks. Available functions vary by language; distinguishing voices does not identify people.
The model card lists about 34 billion total parameters, with 4.3 billion active, and a reference deployment on an 80 GB H100 GPU. Compressed versions are available. OpenMDW-1.1 permits commercial and non-commercial use subject to its terms. This processes files rather than live streams; Nvidia's benchmark claims do not replace evaluation on a user's own recordings.