- Nemotron 3.5 Transcribe has 34 billion total parameters, with about 4.3 billion active per token, and is licensed under OpenMDW-1.1. <sup>[1]</sup>
- The model handles recordings of up to one hour, supports translation between English and 31 languages, and runs locally on hardware such as an 80GB H100. <sup>[1]</sup>
- Independent comparisons show strong accuracy, but the model is a batch system rather than a live-streaming model and is much larger than some streaming alternatives. <sup>[2]</sup><sup>[3]</sup>
Nvidia has released Nemotron 3.5 Transcribe, an open-weight speech model that combines transcription, translation, speaker attribution and audio question answering in one checkpoint. Nvidia’s model card lists 34 billion total parameters, about 4.3 billion active parameters, a 65,536-token context window and support for recordings up to one hour. [1]
The model supports automatic speech recognition in 40 languages and translation to or from English for 31 of them. It can label speakers in supported tasks, produce word timestamps and process audio offline, but Nvidia says internal chunking does not make it a real-time streaming system. The published deployment guidance calls for one 80GB H100 GPU, while supported alternatives include Nvidia’s A100, Blackwell systems, RTX 5090 and Jetson Thor. [1]
Data and benchmark limits #
Nvidia released Granary v2 at the same time. The dataset repository contains manifests covering 463,062 hours across 48 languages and 16 source entries; it distributes labels and instruction pairs rather than the underlying audio, which users must obtain under each source corpus’s terms. Most materials use the OpenMDW-1.1 license, while the People_Speech subset uses CC BY-SA 4.0. [4]
The model card reports leading open-weight results on public word-error-rate benchmarks, including 4.62% mean WER across seven languages and 10.98% on a long-form track. Independent comparisons add context: Ocular AI compared the batch model with three sub-1B streaming systems, while XenoSpectrum reported lower WER from ElevenLabs Scribe v2 on several commercial-service comparisons. Those tests measure different operating constraints, so they do not establish a single overall ranking. [1][2][3]
Companies mentioned #
Further sources #
[\[1\] Nvidia’s model card lists the model’s parameter count, language coverage, licen… ↗](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B)
[\[2\] Ocular AI’s independent evaluation compares Nemotron 3.5 Transcribe with smalle… ↗](https://www.useocular.com/blog/evaluating-nvidia-nemotron-3-5-transcribe-on-converse)
[\[3\] XenoSpectrum reports independent comparisons involving Nemotron 3.5 Transcribe … ↗](https://xenospectrum.com/en/nvidia-nemotron-3-5-transcribe-comparison/)
[\[4\] The Granary v2 dataset page documents the 48-language, 463,062-hour manifest re… ↗](https://huggingface.co/datasets/nvidia/Granary-v2)
The stories that matter, in one email. Free — unsubscribe anytime.