NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt

Why it matters

NVIDIA has released NVIDIA Nemotron 3.5 Transcribe, an open-weight, promptable speech model that unifies transcription, speaker attribution, translation, and audio question answering in a single checkpoint. The model is distributed through Hugging Face and ModelScope, and NVIDIA pairs it with Granary v2, an open multilingual data pipeline plus fine-tuning recipes built on NeMo Speech tools. The pitch is flexibility: developers can steer output with natural-language instructions, adapt the model to domain vocabulary such as medical terminology, and choose where audio is processed, with quantized versions planned for local hardware and Blackwell data-center GPUs. The model ranks first among open-weight entries on the Hugging Face Open ASR Leaderboard for English, sits on Artificial Analysis's accuracy-versus-active-parameter frontier thanks to a hybrid Mamba-Transformer mixture-of-experts design, and posts the lowest speaker-attribution error on David AI's multilingual diarization leaderboard. A medical fine-tuning case study showed large relative word-error-rate gains over prompting alone. For developers, the story matters because a single adaptable open model now covers tasks that previously required separate systems, with training recipes to customize it.