Skip to content
NewA calmer, faster ZeroHour
StoryModel releaseinfo

Nvidia releases open Nemotron 3.5 Transcribe and Granary v2

4 sources · 4 articlesFirst reported Updated ()2
What's new

October 29: MarkTechPost specifies a mixture-of-experts design with about 4 billion of 30 billion parameters active. The Decoder adds the 5.33 percent English error rate, medical fine-tuning result, and language limits.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

Nvidia’s open Nemotron 3.5 Transcribe handles recorded multilingual speech, not live streams, alongside Granary v2.

Nvidia released Nemotron 3.5 Transcribe, a free open-weight speech model, and Granary v2, an open speech dataset. The model follows text instructions for transcription, translation, speaker identification, summaries, and questions about a recording. It transcribes up to 40 languages, translates 31 to or from English, identifies speakers in 13, handles up to one hour and eight speakers, and does not support live streams. Nvidia says the 30-billion-parameter mixture-of-experts model keeps about 4 billion parameters active, runs on a single GPU or chip from Jetson AGX Thor and DGX Spark to Blackwell systems, and posts a 5.33 percent average English word error rate on Hugging Face’s Open ASR leaderboard. Granary v2 is about 844,000 hours in 48 languages; developers can adapt the model with Granary and NeMo Speech tools.

Key facts
  • Transcribes 40 languages and translates 31 to or from English.
  • Speaker identification covers 13 languages; up to eight speakers and one hour.
  • About 4 billion of 30 billion MoE parameters stay active.
  • Nvidia cites a 5.33 percent average English word error rate.
  • Medical fine-tuning cut one benchmark’s error rate from 30.8 to 11.8 percent.
  • Granary v2 contains about 844,000 hours across 48 languages.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

Timeline

dated events in the reporting · oldest first
  1. Oct 28, 2026Nvidia releases Nemotron 3.5 Transcribe and Granary v2. · TechCrunch

What each outlet adds

beyond the shared facts
  • TechCrunchNo live streams; NeMo adaptation; targets meeting and medical apps.
  • The Decoder5.33 percent WER, medical fine-tuning result, and English-only audio QA.
  • MarkTechPostExplicitly calls the model a mixture-of-experts design.
Vendors
NVIDIANvidia
Products
Granary v2GranaryNeMo SpeechDGX SparkJetson AGX Thor
Organizations
Hugging FaceArtificial AnalysisMistralElevenLabsSalesforce
AI models
Nemotron 3.5 TranscribeVoxtral SmallNemotron 3 DiarizationNemotron 3 Nano OmniVoxtral 3ParakeetLauna-OmniGranary v2

Coverage

oldest first · one row per article · UTC
  1. · 2d ago Hugging Face BlogFirst report

    NVIDIA announces Nemotron 3.5 Transcribe and Granary v2

    The post title describes promptable, adaptable multilingual transcription; the feed had no further details.

  2. · 2d ago TechCrunch · AI

    Nvidia releases open-weight Nemotron 3.5 Transcribe speech model

    Nvidia also shipped Granary version 2 and says the model handles recorded audio only, not live streams.

  3. · 2d ago The Decoder

    Nvidia releases open Nemotron 3.5 Transcribe speech model

    Nvidia says the 30-billion-parameter model keeps about 4 billion active and tops open-weight English ASR models.

  4. · 1d ago MarkTechPost

    NVIDIA releases open Nemotron 3.5 Transcribe, a 30B MoE speech model

    The open model transcribes 40 languages while activating about 4 billion of its 30 billion parameters.