Four billion active parameters, tied for first
Nvidia released Nemotron 3.5 Transcribe on 28 October alongside Granary v2, the dataset it was trained on. The model is about 30 billion total parameters with roughly 4 billion active per token, pairing a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model.
On Artificial Analysis’s AA-WER it reaches 2.8%, tied for first among open-weight models — against a model with 24 billion active parameters, six times as many. On the Hugging Face Open ASR leaderboard Nvidia reports a 5.33% average word error rate in English, first among open-weight entries, at an RTFx of 462.09; 4.62% mean WER across seven languages on the multilingual split; and 10.98% on long-form audio, the lowest in the comparison it ran.
One checkpoint does transcription, speaker attribution, translation and audio question answering. It transcribes 40 languages, translates 31 to or from English, attributes speakers in 13, handles recordings up to an hour and labels up to eight speakers. Audio question answering is trained in English only, and the model is built for recorded audio rather than live streaming.
Diarisation, and a dataset with a number on it
On David AI’s diarisation evaluation across 21 languages and 14 systems, Nemotron 3.5 Transcribe posts the lowest diarisation error rate at 18.26%, narrowly ahead of Nvidia’s own Nemotron 3 Diarization at 18.38%, with the rest of the field between 22.85% and 37.55%. On VoiceArena’s English Diarization-Bench it leads 14 systems at 32.4% cpWER, against 34.2% for Multitalker Parakeet with Nemotron 3 Diarization and a field running from 34.7% to 71.9%.
Granary v2, the training corpus, is the part with the hardest number attached: 463,062 hours across 48 languages drawn from 16 source corpora, against 25 languages in version 1. The manifests are public on Hugging Face and ModelScope, but the source audio has to be downloaded separately under each corpus’s own licence. Nvidia’s filtering stage cut VoxPopuli from 354,000 hours to 111,000 across 22 European languages.
Fourteen GPU-hours of medical fine-tuning
The adaptation case study is the most concrete result in the post. On an Eka Care benchmark of 3,619 clips of Indian doctors and medical students, the model started at 30.8% WER on the full test set, fell to 23.8% with a better prompt naming English and the Latin alphabet, and reached 11.8% after fine-tuning — about 62% fewer errors. On the 2,764 held-out clips, WER fell from 22.44% to 12.17%, a 45.8% relative improvement, beating an open medical specialist model at 13.79% and a general ASR model at 14.92%. LibriSpeech test-clean rose slightly, from 1.54% to 1.69%, which is the trade-off.
The training ran on about 153 hours of synthetic speech plus roughly 92 hours of public accented and medical recordings, and took about 14 GPU-hours on a single Nvidia GB300.
What the post does not say
It calls the model open-weight and never names a licence — the post points readers to the model card instead. That matters more than it sounds: “open-weight” covers everything from Apache 2.0 to a bespoke agreement with usage restrictions, and the difference decides whether the model can be used commercially.
The other thing to watch is the quantised build. Nvidia says NVFP4 versions are coming and cites 100 hours of audio transcribed in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU for the forthcoming version at high concurrency on 30-second clips. That figure is for software that has not shipped yet.
