NVIDIA released Nemotron Transcribe weights and the Granary v2 speech-data collection. The model activates about 4 billion of its 30 billion parameters and supports transcription in 40 languages plus translation between English and 31 languages. Speaker diarization covers 13 languages, and the model handles long recordings rather than real-time streaming. The public dataset is a separate collection, not the complete training mixture. Quantized models and some lightweight runtime support are still forthcoming and should be distinguished from the currently available vLLM path.