NVIDIA's Open Nemotron 3.5 Transcribe Puts Transcripts, Speaker Labels and Translation in One Model, for Recorded Audio Only
NVIDIA released Nemotron 3.5 Transcribe on October 28, an open-weight 30B-parameter model with about 4B active that transcribes 40 languages, translates 31 and attributes speakers in 13, for recordings up to an hour. It also released the Granary v2 speech dataset. NVIDIA's leaderboard claims, including first among open models on Open ASR English, are its own; the model does not support streaming.
NVIDIA's new speech model is aimed at a job that usually takes several systems: turning a long recording into a transcript that says who spoke, and translating it. Nemotron 3.5 Transcribe, released with open weights on October 28, handles transcription in 40 languages, translation for 31 and speaker attribution for 13 in one model, according to NVIDIA. What it does not do is live audio: NVIDIA says it is built for recorded files, not streaming.
What NVIDIA released#
The model has 30 billion total parameters with about 4 billion active per token, using a hybrid Mamba-Transformer mixture-of-experts design, NVIDIA says. It takes recordings of up to an hour with up to eight speakers. NVIDIA also released Granary v2, a speech dataset it describes as 463,062 hours across 48 languages, and said quantized versions of the model are coming.
TechCrunch framed the release as part of a wider push into voice AI. Mistral released its own Voxtral 3 speech models last week, and NVIDIA's previous Nemotron 3.5 release was the Super VL multimodal model.
The leaderboard claims#
NVIDIA reports a 5.33% average word error rate on the Hugging Face Open ASR leaderboard for English, first among open-weight models, and a 2.8% score on Artificial Analysis's AA-WER, tied for first among open-weight models. It also reports an 18.26% diarization error rate on a David AI benchmark and 32.4% on VoiceArena. NVIDIA says a medical fine-tune cut word error rate by 45.8% relative to the base model on its tests.
What the model card adds#
The model card on Hugging Face fills in two things the announcement left open. The weights are released under OpenMDW-1.1, and the card says the model "is ready for commercial or non-commercial use." FP8 and NVFP4 versions, the quantized releases NVIDIA described as coming, are now listed on NVIDIA's Hugging Face page alongside the full model.
The card's own headline numbers do not match the announcement's. It reports a mean word error rate of 4.71% on nine public sets of the Open ASR leaderboard, captured on October 12, against 4.94% for Qwen3-ASR-1.7B, the next best open-weight model it lists; the announcement gives 5.33%. Neither source explains the gap, which may come from a different set of test sets or a later run. The card also says the fastest serving setup it describes, NVIDIA's optimised vLLM implementation inside its Nemotron Transcribe NIM, is "not yet part of the open-source release." Developers taking the open weights will not get those speeds out of the box.
Limits#
- All benchmark and leaderboard figures above are as reported by NVIDIA, in its announcement or on the model card, and have not been checked against the live leaderboards.
- No streaming or real-time support; NVIDIA says the model is for recorded audio.
- OpenMDW-1.1's full terms are on the licence page linked from the card; this article does not summarise them.
Sources#
- Hugging Face: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card
- Hugging Face: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B-FP8
- NVIDIA on Hugging Face: Nemotron 3.5 Transcribe and Granary v2 (October 28, 2026)
- Hugging Face: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model card
- TechCrunch: Voice AI in vogue, Nvidia releases new open transcription model (October 28, 2026)