NVIDIA Nemotron 3.5 Transcribe
An open-weight model for prompted recorded-audio transcription, translation, speaker attribution and English audio questions.
What happened lately
NVIDIA Nemotron 3.5 Transcribe timeline
NVIDIA Nemotron 3.5 Transcribe is a prompted speech model released on October 28, 2026. NVIDIA describes transcription in 40 languages, translation to and from English for 31 languages, speaker attribution in 13 languages and English audio question answering. Those task-specific limits differ; they do not imply that every task works across every language.
The model card describes recorded-audio processing rather than real-time streaming, with support for recordings up to one hour. It lists BF16, FP8 and NVFP4 variants under OpenMDW-1.1. The reference quickstart uses one H100 with 80 GB of memory, vLLM 0.32.0 or later and the latest NeMo Speech code. NVIDIA’s fastest NIM serving claims use optimizations not yet present in its open-source integration at announcement.