ModelRadar
/

Nemotron-3.5-Transcribe

NVIDIAπŸ‡ΊπŸ‡Έ USMajor labΒ· Oct 7, 2026

Audioopen weights

Nemotron 3.5 Transcribe is an open-weight, offline speech and audio model for transcription in 40 languages, with support for recordings up to one hour and a 65,536-token context.

From the model card

License
other
Modalities
audio β†’ text

1 variant

30B A4B30B-A4BHF

Sources

HuggingFacePaper

Provenance

Detected
Oct 29, 2026
First source
huggingface
Description
From the model card
Editorially reviewed
β€”