NVIDIA released Nemotron 3.5 Transcribe on Wednesday, October 28, 2026, an open-weight speech and audio model that transcribes recordings in 40 languages. According to its Hugging Face model card, it handles recordings of up to one hour, a 65,536-token context, and may be used for commercial and non-commercial purposes under the OpenMDW-1.1 licence.
The model has 34 billion parameters in total, with about 4.3 billion active at inference, including the audio encoder, in a hybrid Mamba-Transformer mixture-of-experts architecture. It transcribes audio that has already been recorded — it is not a live streaming ASR service, an important distinction from the family's streaming ASR model.
According to the model card, the reference deployment uses one 80 GB H100 GPU; TechCrunch reports that Nvidia says the model can run on a single chip, from the Jetson AGX Thor robotics controller to Blackwell datacentre chips.
What the model does
Beyond verbatim or readable transcription, it offers speech translation, speaker-attributed transcription, language identification, word timestamps and audio question answering. According to TechCrunch, it can tell up to eight speakers apart and translate 31 languages to and from English, with domain vocabulary customisation through NeMo Speech.
What it is for
NVIDIA cites meetings, medical transcription, media processing and customer-service analytics among its uses. The company's own benchmark results claim the best open-weight result on the English Open ASR Leaderboard, with a mean error rate of 4.71% — supplier figures, not yet independently verified. NVIDIA also released version 2 of its Granary speech dataset alongside the model.
In the same Nemotron 3.5 family, Geekiki previously covered the 4 GB Bolt E4B for local agents — Transcribe is the transcription-focused member, not a real-time voice agent.
Image credit: NVIDIA, via the model page on Hugging Face.
Sources: Hugging Face — NVIDIA Nemotron 3.5 Transcribe; TechCrunch.
Transparency: this article was produced with the help of artificial intelligence, cross-checked against public posts on X and internet sources. We recommend consulting the original sources for more details.
Muse
