NVIDIA AI introduced Nemotron 3.5 Transcribe, "a promptable 30B MoE model with open weights and ~4B active parameters" that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint. The weights are on Hugging Face, with a live demo and a domain-adaptation guide linked from the announcement.
What it does
Instead of a fixed transcriber, the model is steered with a text prompt alongside the audio. NVIDIA's examples are adding speaker labels to a transcript, translating it into Spanish and then Japanese, and finding and redacting personally identifiable information. The demo shows color-coded captions with speaker labels.
- Length and speakers: recordings up to an hour, with up to eight speakers labelled.
- Languages: 40 for transcription, with support varying by task, so translation and speaker attribution do not cover every language.
- Context: the model card lists a 65,536-token context and BF16, FP8 and NVFP4 precisions.
- Size: the announcement says 30B with about 4B active. The model card counts the audio encoder and lists 34B total with 4.3B active.
- License: OpenMDW-1.1, according to coverage of the release.
The numbers
NVIDIA says Nemotron 3.5 Transcribe ranks first among models with open weights on the Hugging Face Open ASR leaderboard for English, at a 5.33% average word error rate. Coverage also cites Artificial Analysis, where it ties for first among open-weight models at 2.8% AA-WER with Mistral's Voxtral Small, which has about 24B parameters (unrounded, 2.81% against 2.77%), so the same accuracy comes from about a sixth of the active parameters. Both figures are as reported by NVIDIA and coverage of the release.
Granary v2
NVIDIA released the model with Granary v2, an open speech dataset whose manifests cover 463,062 hours of audio in 48 languages, according to coverage of the launch.
It follows NVIDIA's Nemotron 3.5 Super VL earlier this month, and joins a crowded transcription field that includes Google's Gemini 3.5 Transcribe and Block's speech-only Dessa.