NVIDIA releases Nemotron 3.5 Transcribe, a promptable speech model

NVIDIA published Nemotron 3.5 Transcribe on 28 October 2026 under OpenMDW-1.1: 40 languages, speaker labels, translation and prompt-controlled transcript style, for recordings up to an hour.

  • New release
  • 2 min read
  • Published
  • Updated

NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026 under the permissive OpenMDW-1.1 license. Along with the model, it published a fine-tuning toolkit in NeMo Speech and version 2 of its Granary speech dataset.

What is new

  • One model, many jobs. A speech encoder and a diarization model, which works out who spoke when, feed a mixture-of-experts language model. It handles transcription, speech translation, speaker-attributed transcripts, language identification and questions about audio, and it can return word-level timestamps.
  • Size. The model card lists about 34 billion parameters in total and 4.3 billion active, including the audio encoder.
  • Coverage. It supports 40 languages, with translation to or from English in 31 of them and speaker attribution in 13, on recordings of up to one hour with up to eight speakers.
  • Steerable style. You describe the transcript you want in a text instruction, for example verbatim text or written-form numbers, or you supply a list of names and terms to expect.
  • Offline only. It processes finished recordings. It is not a live-streaming recognizer.

NVIDIA reports a mean word error rate of 4.71% on the English Open ASR Leaderboard and 4.62% on the multilingual one. These are the developer’s own measurements, so test the model on your own audio.

Running it

The BF16 weights are about 68 GB, the FP8 version about 35 GB and the NVFP4 version about 22 GB. The model card lists a single 80 GB H100 as its deployment target, and GGUF files are available for NVIDIA’s NeMo-Speech.cpp. That makes it much heavier than Whisper large-v3-turbo, which needs roughly 6 GB. Our side-by-side comparison shows when the extra size is worth it, and the speech models page lists the alternatives.

License

OpenMDW-1.1 is a short permissive license written for machine-learning models. It lets you use, modify, share and sell the model and places no restrictions on its outputs. You must keep the license text and the origin notices when you redistribute it, and the Open Source Initiative has not approved the license. AI licenses explained covers how it compares with Apache 2.0 and MIT.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU