On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model that packs transcription, speaker diarization, translation, and recorded-audio question answering into a single model. Once developers download the weights from channels such as Hugging Face, they can use natural-language instructions to turn the same recording into a verbatim transcript, a cleaned-up text, a summary, or answers to questions, without wiring up a separate model for each task.
One model, four jobs
Nemotron 3.5 Transcribe supports transcription in 40 languages, translation between 31 languages and English, speaker separation for up to eight speakers in 13 languages, recordings up to one hour long, and word-level timestamps. Change one instruction and the same recording comes out in a different form: a verbatim transcript, a cleaned-up text, a meeting summary, or answers about what was said. Instructions can also specify terminology and number formatting, which matters in jargon-heavy fields such as healthcare and finance.
It runs on one GPU, but only for recordings
The model is a 30-billion-parameter mixture-of-experts design that activates about 4 billion parameters per inference step. NVIDIA says it can run on a single GPU, from the Jetson AGX Thor and DGX Spark to Blackwell data-center chips; at NVFP4 precision, one RTX PRO 6000 processed 100 hours of audio in under three minutes, according to the company's own figures. The limits are equally clear: it handles recordings only, not live transcription, and its audio question answering was trained mainly in English, so other languages need testing first.
The dataset is open too, but treat benchmark claims with care
Released alongside the model, Granary v2 covers 463,062 hours of speech across 48 languages through an open dataset and curation pipeline, up from 25 languages in the previous version. The weights use the OpenMDW-1.1 license, which permits commercial use. NVIDIA says the model leads open-weight models on Hugging Face's Open ASR Leaderboard, a vendor claim: different leaderboards use different methods, so selection should still be tested on your own recordings. Voice releases have been coming thick and fast, and Nemotron's difference is not parameter count but engineering: one checkpoint plus customization. For teams with fixed workflows, such as meeting notes or medical transcription, the cost of switching models often matters more than a few extra leaderboard points.
