NVIDIA releases Nemotron 3.5 Transcribe, an open speech model for 40 languages, and Granary v2

What NVIDIA released

NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026, announced in a post on the Hugging Face blog. It is an open-weight, promptable speech model that combines transcription, speaker attribution, translation and audio question answering in one checkpoint. It transcribes 40 languages, translates 31 of them to or from English, and labels speakers in 13. It handles recordings up to an hour long with up to eight speakers.

The model pairs a speech encoder with a hybrid Mamba-Transformer mixture-of-experts language model, so only about 4B of its 30B parameters are active per token. Weights are on Hugging Face and ModelScope under the OpenMDW-1.1 licence. Developers can steer output with plain instructions, such as asking for speaker labels or giving a list of medical terms to watch for. NVIDIA notes it is built for recorded audio, not live streaming.

Benchmark claims

NVIDIA says Transcribe ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English, at 5.33% average word error rate, and reports 4.62% across seven languages and 10.98% on long-form audio. On Artificial Analysis's speech-to-text index it ties for first among open models at 2.8% word error rate, with about one-sixth the active parameters of the model it ties. NVIDIA also cites the lowest diarization error rate, 18.26% across 21 languages, on David AI's leaderboard, and first place in VoiceArena's 14-system comparison.

For throughput, NVIDIA says a forthcoming NVFP4 quantized version processed 100 hours of audio in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU. Quantized versions for local use on RTX Spark, DGX Spark and Jetson AGX Thor are promised but not yet out.

Granary v2 and fine-tuning

NVIDIA also published Granary v2, the data curation pipeline and manifests behind the model's training. The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in the first Granary. The manifests carry verbatim and cleaned transcripts, translations, labels and quality flags. Source audio must be downloaded separately under each corpus's own licence.

A medical case study shows how far adaptation goes. On held-out speakers from Eka Care's Indian medical speech benchmark, fine-tuning cut word error rate from 22.44% to 12.17%, a 45.8% relative reduction, using about 14 GPU-hours on one GB300. NVIDIA ships the recipe as NeMo Speech tutorials and an agent skill that lets a coding agent run the workflow.

What it means and what to watch

Transcribe adds a strong open option for meeting, call and dictation transcription where teams want to keep audio on their own servers. It follows NVIDIA's Nemotron 3.5 Bolt release for local agents. Watch for the quantized builds, independent leaderboard checks and the terms of the OpenMDW licence for commercial use. More open model news is on the Open Source AI hub.

Questions people ask

What licence does Nemotron 3.5 Transcribe use?

The weights are released under the OpenMDW-1.1 licence, according to the Hugging Face model card.

How many languages does Nemotron 3.5 Transcribe support?

It transcribes 40 languages, translates 31 of them to or from English and attributes speakers in 13. Audio question answering is trained in English.

What is Granary v2?

NVIDIA's open speech data pipeline and manifests, covering 463,062 hours in 48 languages from 16 corpora, used to train Nemotron 3.5 Transcribe.

Sources

1. NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt (NVIDIA on Hugging Face, 28 Oct 2026)
2. NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card (NVIDIA, 28 Oct 2026)
3. Granary v2 dataset (NVIDIA, 28 Oct 2026)