NVIDIA has released Nemotron 3.5 Transcribe, an open-weight, promptable ASR model handling transcription, speaker diarization, translation, and audio Q&A in a single 30B-parameter checkpoint with a hybrid Mamba–Transformer MoE architecture activating only ~4B parameters per token. It supports 40 languages for transcription, 31 for translation to/from English, and 13 for speaker attribution (up to eight speakers, hour-long recordings). On the Hugging Face Open ASR Leaderboard, it ranks first among open-weight models for English (5.33% avg WER), multilingual (4.62% mean WER across seven languages), and long-form (10.98% WER). Artificial Analysis reports 2.8% AA-WER (tying the best open model at one-sixth the active parameters) and 1.8% WER on the held-out AA-AgentTalk set. For diarization, it achieves the lowest DER (18.26%) on David AI’s 21-language benchmark and leads VoiceArena’s 14-system English comparison (32.4% cpWER, 8.32% deletions vs. ~31% median). High-throughput inference processes 100 hours in under three minutes on one RTX PRO 6000 Blackwell (NVFP4 quantized). The release includes Granary v2, a reproducible NeMo Curator pipeline and 463k-hour multilingual dataset (48 languages, 16 corpora) producing three transcript forms (verbatim, ITN with/without disfluencies) plus captions, Q&A pairs, and translations. A medical fine-tuning case study on Eka Care’s Indian English benchmark (3,619 clips) cut held-out WER from 22.44% to 12.17% (45.8% relative reduction) using ~153h synthetic speech, 29h accented English, and 63h medical audio, with LoRA adaptation completed in ~14 GPU-hours on one GB300. Deployment is supported via vLLM, SGLang-Omni, and hosted endpoints (Baseten, DeepInfra, Simplismart), with NeMo-Speech.cpp and quantized versions forthcoming for local RTX/Jetson deployment.