The NVIDIA/Hugging Face article, model card and dataset page jointly confirm weights, task and language scope, data scale, licenses and evidence limits.
NVIDIA opens Nemotron 3.5 Transcribe and Granary v2
2026-10-28Verified announcementTRIAL
NVIDIA released open-weight Nemotron 3.5 Transcribe for multilingual transcription, speaker attribution, translation and audio question answering. The 30B mixture-of-experts model activates about 4B parameters, handles recordings up to one hour and supports task-dependent language sets. Granary v2 publishes a reusable curation pipeline and manifests spanning 463,062 hours, while source audio remains under each corpus's original license.
What this means for you
For speech and meeting-intelligence teams, one checkpoint now covers several pipeline stages with public weights and adaptation recipes. Vendor and leaderboard results still need workload-specific validation, especially names, speaker changes, long recordings and license lineage.
Evidence & related material
Technical details & scores
| Model | 30B total, about 4B active parameters; OpenMDW 1.1 |
|---|---|
| Tasks | transcription, speaker attribution, translation, timestamps and English-trained audio question answering |
| Language Scope | 40 transcription languages; 31 translation languages to/from English; 13 diarization languages |
| Audio Limit | recordings up to one hour; up to eight speakers |
| Dataset | Granary v2 manifests cover 463,062 hours, 48 languages and 16 source corpora |
| License Limit | download source audio separately under each corpus's original license |
| Evidence Limit | headline model results are NVIDIA-reported; public leaderboards cover selected tasks and conditions |
| Time Basis | Hugging Face RSS published 2026-10-28T14:58:17Z; model card modified 2026-10-28T13:44:41Z |
Technical impact 4
Engineering value 4
Adoption signal 2
Maturity 3
Verification cost 3
Risk 3