NVIDIA releases Nemotron 3.5 Transcribe, an open-weight 30B-A4B speech model for transcription, diarization and translation in 40 languages, with the Granary v2 dataset
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 29 October 2026
Your AI and this story
- GPT-6 Astra181 days after its cutoff
- Claude Opus 5.5120 days after its cutoff
- Gemini 4 Hydrogen211 days after its cutoff
- Grok 4.7150 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 120 days before it.
Key facts
- Weights on Hugging Face under the OpenMDW-1.1 licence (BF16, FP8, NVFP4), plus NIM/build.nvidia.com, NGC and NeMo Speech on GitHub, all dated Oct 28; served with vLLM 0.32.0
- Architecture: ASR encoder plus Nemotron-3-Diarization speaker encoder feeding the hybrid Mamba-Transformer MoE Nemotron 3.5 Lightning 30B-A3B; CTC timestamp head; native multi-token prediction and an optional DFlash2 draft model; 65,536-token context
- Vendor benchmarks (captured Oct 12): English Open ASR mean WER 4.71% vs Qwen3-ASR-1.7B 4.94%; long-form 10.98% vs Parakeet-TDT-0.6B-v3 13.37%; multilingual Open ASR (7 languages) 4.62% vs 6.49%; multi-speaker mean cpWER 15.80% vs 16.83% for DiCoW v3.3 + DiariZen
- Throughput (NVIDIA): up to 2,238 seconds of audio per second on one RTX PRO 6000 Blackwell Server Edition (NVFP4 with DFlash2); runs from Jetson Thor and DGX Spark to data-centre GPUs
- Limits: offline model, not real-time streaming; audio understanding and reasoning evaluated only in English; recordings up to 1 hour per the model card (SiliconANGLE’s ‘answering questions about 100 hours of audio’ is not in the card; its ‘under three minutes for 100 hours’ is a batch-throughput figure)
- Training data: ~1.72M unique hours (6.53M with repeats), ~372M audio examples; Granary v2 public manifest 463,062 hours across 48 languages and 16 corpora (Granary v1: 25 languages); includes data licensed from David AI
What happened
NVIDIA folded the usual speech pipeline (voice activity detection, transcription, diarization, translation, question answering) into one open model driven by prompts, built on its Nemotron 3.5 Lightning language model. The release includes the training-data pipeline (Granary v2) so others can fine-tune it.
Why it matters
It is the strongest open-weights transcription model by NVIDIA’s own measurements on public leaderboard sets, and its permissive licence lets companies run meeting and call transcription on their own GPUs. Independent leaderboard results were not yet available on Oct 29.
Sources
5 sources from 4 sites. Numbers match the chips in the text.
5 sources: 3 primary, 2 press
Primary
- Hugging Face: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B model cardhuggingface.co, official
- NVIDIA: Nemotron 3.5 Transcribe technical report (PDF)research.nvidia.com, paper
- Hugging Face: Granary v2 datasethuggingface.co, code
Press
Changes
- Filed. Model file nemotron-3-5-transcribe added (news agent)