Nemotron 3.5 Transcribe
NVIDIA released Nemotron 3.5 Transcribe, an open-weight promptable model that transcribes 40 languages and handles recordings up to an hour long, alongside the Granary v2 dataset.
One checkpoint covers transcription, speaker attribution, translation and audio question answering. It translates 31 languages to or from English, labels up to eight speakers and attributes speakers in 13 languages. Developers can steer output with natural language prompts and adapt it to domain vocabulary and accents. It works on recorded audio, not live streams.
- Date
- Wednesday 28 October 2026
- Lab
- NVIDIA
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Languages transcribed | 40 31 translate to or from English, speaker attribution in 13 | company |
| Parameters | 30 billion about 4B active parameters per the NVIDIA blog; runs on a single chip, from Jetson AGX Thor and DGX Spark to Blackwell | company |
| Max recording length | 1 hour up to eight speakers labeled | company |
| Hugging Face Open ASR Leaderboard (English) | first among open-weight models 5.33% average WER; multilingual (4.62%) and long-form (10.98%) results are lowest only within the model card's own comparison | company |
Available on Hugging Face and ModelScope; the license was not stated on the pages read.
Sources
- huggingface.co/blog/nvidia/nemotron-transcribe-granary
- techcrunch.com/2026/10/28/voice-ai-in-vogue-nvidia-releases-new-open-transcription-model/
This record was checked and corrected against its sources on 28 October 2026. How we check