NVIDIA open-sources Nemotron 3.5 Transcribe: 30B MoE ASR model with ~4B active params
NVIDIAAI · x · 2026-10-28
NVIDIA released Nemotron 3.5 Transcribe, an open-weight, promptable 30B MoE model (4B active params) that unifies multilingual transcription, speaker diarization, translation and audio Q&A in one checkpoint: 40 transcription languages, 31 translation languages, speaker attribution in 13, up to 1-hour recordings and 8 speakers, plus natural-language control over output (speaker labels, translation, PII redaction).
Key results:
- #1 open-weight model on Hugging Face Open ASR English leaderboard, 5.33% average WER
- 2.81% WER on Artificial Analysis, nearly matching Voxtral Small (2.77%) with 1/6 the active params
- #1 in VoiceArena's 14-system diarization+ASR comparison and David AI's multilingual diarization benchmark (DER 18.26)
NVIDIA also open-sourced the Granary v2 data pipeline and NeMo fine-tuning recipes: in a medical ASR case study, fine-tuning cut WER by 45.8% relative to the best prompt. Available on Hugging Face and ModelScope.
Related event: NVIDIA Open-Sources Nemotron 3.5 Transcribe, Topping Open ASR Leaderboards(7 posts)→
More from Models
- Cryptographer points abliterated GLM at Telegram's secret-chat crypto — and it 'killed itself' — matthew_d_green · 2026-10-29
- Qoder offers DeepSeek-V4.1-Flash-1017 free through Oct 31 with zero credit cost — Aiden_Tech_Ai · 2026-10-29
- Top reasoning effort costs up to 42x more on tasks low effort already solves — zainhas · 2026-10-29
- RealityBench: 35 VLMs tested on physical work understanding, measuring stays a weak spot — arnabdotorg · 2026-10-28
- Permissive flagship open-model licenses fell from 75% to 35% in months — bookwormengr · 2026-10-28
- Anthropic Gets 80% of Revenue From API Pricing; DeepSeek Costs 20x Less — bindureddy · 2026-10-28