NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt
NVIDIA released an open-weight speech model alongside a multilingual dataset and curation pipeline.
TL;DR
- NVIDIA released Nemotron 3.5 Transcribe and Granary v2 on 2026-10-28.
- The model is designed to transcribe 40 languages and translate 31 to or from English.
- The released dataset manifest covers 463,062 hours across 48 languages.
NVIDIA released Nemotron 3.5 Transcribe, an open-weight model for recorded speech. The company also made Granary v2 available as a multilingual data resource and curation pipeline. [1] [2]
NVIDIA says the same model can transcribe 40 languages, translate 31 to or from English, and identify speakers in 13. It says the model handles recordings up to 1 hour long and can label up to 8 speakers. [1]
The public Granary v2 manifest covers 463,062 hours across 48 languages. NVIDIA specifies that the manifest describes the dataset release, not the model’s complete training mixture. [1] [3]
TechCrunch and The Next Web reported that the model works on recorded audio rather than live streams. NVIDIA’s benchmark and adaptation results are company-reported; their performance across real deployments remains open. [1] [2] [3]
Why it matters
An open-weight release gives developers another way to build speech applications while retaining more control over deployment. Its practical value will depend on licensing, hardware and performance across languages and recording conditions.
Editor's note
NVIDIA authors published the primary source on Hugging Face; TechCrunch and The Next Web reported the release independently. Capability and performance claims are attributed to NVIDIA rather than treated as independently verified results.