NVIDIA opens Nemotron 3.5 Transcribe with weights, data and recipes

NVIDIA describes Nemotron 3.5 Transcribe as open and customizable, explicitly including the model’s weights, data and recipes.[3] The model is identified on Hugging Face as NVIDIA-Nemotron-3.5-Transcribe-30B-A4B.[1] Its model card lists 34 billion total parameters and 4.3 billion active parameters, including the audio encoder.[1] NVIDIA says a single checkpoint brings together transcription, speaker attribution, translation and audio question answering.[2] For teams evaluating it, the combination means these tasks can be assessed as part of one model rather than treated as separate listed capabilities.

NVIDIA says the model transcribes speech in 40 languages, translates 31 languages to or from English, and attributes speakers in 13 languages.[2] It can process recordings up to one hour long and label as many as eight speakers.[2] The model card also says users can provide natural-language instructions and adapt recognition to domain vocabulary and accents.[2] One important constraint is that it processes recordings offline; NVIDIA does not present it as a real-time streaming speech-recognition service.[1] Teams should therefore check whether their workflow can accept recorded audio rather than live transcription.

The model is available on Hugging Face and ModelScope, and NVIDIA says it is ready for commercial or non-commercial use.[2][1] NVIDIA reports a 5.33% average word error rate and first place among open-weight models on Hugging Face’s English Open ASR Leaderboard.[3] Its model card reports a 4.62% mean word error rate across seven languages in its Open ASR comparison.[2] NVIDIA says the model can run on a single GPU, ranging from Jetson AGX Thor and DGX Spark systems to Blackwell data-center GPUs.[2] On one RTX PRO 6000 Blackwell Server Edition GPU, it reports processing 100 hours of audio in under three minutes under its stated test conditions.[2]

Granary v2 provides multilingual data on Hugging Face and ModelScope, along with a reusable curation pipeline.[2] NVIDIA says its public manifest covers 463,062 hours of audio across 48 languages and 16 source corpora.[2] For teams seeking to adapt recognition beyond prompt-level instructions, NVIDIA says NeMo Speech fine-tuning tools can extend that control to training.[2] In a medical ASR case study, NVIDIA reports that fine-tuning reduced word error rate by 45.8% relative to the best prompt on held-out speakers.[3] That result is a reported case study, not a stated guarantee for other datasets or deployments.