1 source 52m ago
NVIDIA Releases Nemotron 3.5 Transcribe as an Open Model
NVIDIA released the transcription AI Nemotron 3.5 Transcribe on October 28, 2026.
Reporting from 1 source: GIGAZINE.
Why it matters
NVIDIA is shipping a free open transcription model that it rates as the top-performing open entry on the Open ASR Leaderboard's English section and on the Artificial Analysis Word Error Rate Index, which puts a 4.3B-active-parameter model in the same accuracy conversation as much heavier ones.
Key Facts
- NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026.
- Nemotron 3.5 Transcribe has 34 billion total parameters and 4.3 billion active parameters.
- The model transcribes audio files up to one hour long and supports 40 languages, with Japanese rated Production-ready.
- Nemotron 3.5 Transcribe is free under the OpenMDW-1.1 license.
- NVIDIA also released the Granary v2 dataset used to develop the model.
The model has 34 billion total parameters and 4.3 billion active parameters, handles audio files up to one hour long, and supports 40 languages including Japanese, which NVIDIA rates in its Production-ready accuracy tier. The model is free under the OpenMDW-1.1 license, and the Granary v2 training dataset is also out.
The model is a mixture-of-experts design: 34 billion total parameters but 4.3 billion active at a time. NVIDIA's published graph plots active parameter count against error rate and places Nemotron 3.5 Transcribe at the fewest active parameters among models of comparable accuracy.
NVIDIA splits transcription accuracy into three tiers: Leading, Production-ready, and Supported. Japanese lands in Production-ready. The demo transcribes rapid overlapping dialogue from the 1940 film His Girl Friday and separates the speakers while doing it. On the Artificial Analysis Word Error Rate Index for non-streaming transcription, the model shares the top open-model rating with Voxtral Small Transcribe. A quantized version with lower memory use is planned.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google