NVIDIA releases open-weight Nemotron 3.5 Transcribe multilingual speech model and Granary v2 data pipeline

NVIDIA published Nemotron 3.5 Transcribe, an open-weight speech model (about 30B total parameters, roughly 4B active per token), on Hugging Face and ModelScope, alongside the Granary v2 multilingual dataset, a reusable data curation pipeline and NeMo Speech fine-tuning recipes. One model covers transcription in 40 languages, translation for 31 languages to and from English, speaker labelling in 13 languages (up to 8 speakers, recordings up to an hour), word-level timestamps and English audio Q&A. NVIDIA reports a 5.33% average word error rate on the Hugging Face Open ASR Leaderboard (first among open-weight models, English) and 18.26% diarization error on David AI's benchmark. Quantised versions are promised but not yet released.

Teams building transcription, meeting-notes or call-analytics products now have one downloadable model that replaces several separate tools, which could cut cost and keep audio in-house. The dataset and curation pipeline are arguably as useful as the model, since they help others build or adapt speech systems for less common languages. Practical value depends on running costs: a 30B-parameter model still needs substantial GPU memory even with about 4B active, so small teams may wait for the quantised versions. The medical fine-tuning gain (45.8% relative error reduction) suggests domain adaptation works, but it is a single case study.

The leaderboard rankings, the medical case study and the "one model does everything" framing all come from NVIDIA's own post, and no independent replication is cited. "First among open-weight" is a narrow claim: it covers English on one leaderboard, and says nothing about the other 39 languages, noisy audio or real-time latency. Accuracy on the diarization and translation features is far less established than on plain English transcription, and licence terms and hardware needs are not covered in the summary.

1. 01Independent benchmark results on non-English languages, noisy real-world audio and long recordings, from third parties rather than NVIDIA, in the next few weeks
2. 02Release of the promised quantised versions and published memory, speed and cost figures, which will show whether small teams can actually run it
3. 03Download counts, community fine-tunes and named production adopters on Hugging Face, plus the licence terms for commercial use of the model and Granary v2