NVIDIA releases Nemotron 3.5 Transcribe speech model

On October 28, 2026, NVIDIA released the open model Nemotron 3.5 Transcribe (30B-A4B), which combines speech transcription, translation, speaker diarization, and audio question answering in a single checkpoint across 40 languages, including Russian.

Compiled by Sergey KostenchukPublished 2026-10-29Updated 2026-10-29

2026-10-29 Research NVIDIA

🤖 NVIDIA releases Nemotron 3.5 Transcribe speech model

The open model (30B-A4B) combines speech transcription, translation, speaker diarization, and audio question answering: 40 languages, including Russian, translation of 31 languages to/from English, up to 8 speakers, and recordings up to one hour long. 34 billion total parameters, ~4.3 billion active.

🌍 The model topped the open ASR leaderboard on Hugging Face: 5.33% WER on English and 4.62% on the multilingual portion; Artificial Analysis reports a 2.8% WER. The Granary v2 dataset was also released — 463,000 hours of speech in 48 languages for fine-tuning.

👤 Weights are open (OpenMDW-1.1) on Hugging Face and ModelScope: local deployment via vLLM and NeMo Speech, demo available on Hugging Face Spaces. Limitations: audio Q&A is only available in English, and live streaming is not supported.

Source 1: https://huggingface.co/blog/nvidia/nemotron-transcribe-granary Source 2: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B

Sources