• Sources: primary, discussion
  • Summary: Nemotron 3.5 Transcribe transcribes 40 languages, translates 31, attributes speakers in 13 and handles recordings up to an hour long, with about 4B of its 30B parameters active per token. NVIDIA reports 5.33% average WER on the Open ASR English leaderboard, the best among open-weight models. It also released Granary v2 manifests covering 463,062 hours in 48 languages. vLLM and SGLang-Omni support it now, and quantized builds are listed as coming soon.
  • Why it matters: Self-hosted transcription pipelines can replace separate ASR, diarization and translation models with one promptable checkpoint.