NVIDIA releases Nemotron 3.5 Transcribe, an open promptable speech model

NVIDIA released open-weight Nemotron 3.5 Transcribe, which handles transcription, speaker labels, translation and audio Q&A, plus the Granary v2 dataset.

NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026. It is an open-weight model that handles transcription, speaker labels, translation and audio Q&A in one checkpoint. You pick the task with a plain-language prompt, such as "Transcribe this recording and label each speaker." For anyone building meeting-notes or dictation pipelines, that means one self-hosted model where you would otherwise chain several tools or send audio to a third-party API.

NVIDIA's post gives these limits:

It is built for recorded audio, not live streams, a drawback TechCrunch also flags. Alongside the model, NVIDIA released Granary v2, a 463,062-hour dataset, and NeMo Speech fine-tuning recipes. In a medical case study, NVIDIA says fine-tuning cut word error rate from 23.8% with a tuned prompt to 11.8%.

The hardware bar is high. The name says 30B, but MarkTechPost puts the checkpoint at about 34B total parameters, with about 4.3B active. MindStudio observed just over 75 GB of VRAM when running it in BF16 through vLLM. NVIDIA's research page lists FP8, NVFP4 and GGUF checkpoints, but the Hugging Face post calls the NVFP4 version "forthcoming." Check what is actually downloadable before you plan on a smaller card.

On benchmarks, NVIDIA claims first place among open-weight models on the Open ASR Leaderboard. Tech Times notes these figures are NVIDIA-sourced. MarkTechPost reports that commercial APIs still lead, including ElevenLabs Scribe v2 on the multilingual track.

The license is OpenMDW-1.1. MarkTechPost says it allows commercial use, while MindStudio advises reading the terms carefully.

Sources