NVIDIA released Nemotron 3.5 Transcribe on October 28, 2026. Its Hugging Face blog calls it "an open-weight, promptable model that turns recorded speech into useful text." It "brings transcription, speaker attribution (diarization), translation, and audio question answering into one checkpoint." NVIDIA also released Granary v2, the speech dataset pipeline used to prepare its training data. TechCrunch described the model as "a free, open-weight model that transcribes audio in up to 40 languages."

#What can Nemotron 3.5 Transcribe do?

In the blog's words, "it transcribes 40 languages, translates 31 of them to or from English, and attributes speakers in 13." It "handles recordings up to an hour long and labels up to eight speakers." Developers steer it with plain prompts, such as "Transcribe this recording and label each speaker." or "What decisions were made in this meeting?" NVIDIA adds that "Word-level timestamps are also supported."

There are limits NVIDIA states itself:

  • Recorded audio only. "Transcribe is built for recorded audio, not live streaming."
  • Language support varies. "language support differs by task, and audio question answering is trained in English."
  • Speaker labels are not identity. The model card says labels "do not establish real-world identity."
  • Check the output. "As with any speech model, spot-check names, speaker changes, and answers against the recording."

The card lists .wav, .flac, .mp3 and .ogg input.

#How big is it, and how is it licensed?

The model card gives "34B (4.3B active), including the audio encoder." The Hugging Face blog rounds differently, saying "about 4B of its 30B parameters are active per token"; the card's figure includes the roughly 800M-parameter audio encoder. The card describes the architecture as a "hybrid Mamba-Transformer MoE with Transformer ASR encoder and Sortformer speaker encoder."

The license, per the card, is the "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)." The card says "This model is ready for commercial or non-commercial use." The BF16 checkpoint is not gated on Hugging Face. NVIDIA lists no price; "free" is TechCrunch's description.

#What is Granary v2?

The blog says Granary v2 was "Built with NVIDIA NeMo Curator," and that "its reusable pipeline prepared multilingual speech used in Nemotron 3.5 Transcribe’s training, producing aligned transcript variants, task labels, and inspectable quality metadata."

  • Size. "The public manifest covers 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1."
  • Caveat. "These figures describe the dataset release, not Transcribe’s full training mixture or its sampling weights."
  • Outputs. "Granary v2 generates three transcript forms for each recording," plus "short audio captions, code-switching restoration of loanwords, speech-grounded Q&A pairs, contextual ASR biasing terms, and translations with automatic quality scores."
  • No audio. The dataset card says "This repository contains label manifests only." The blog says to "Download source audio separately under each corpus’s original license."

The dataset card gives the license: "This dataset is governed by the OpenMDW License Agreement, version 1.1, except for the People_Speech subset, which is governed by the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)."

#How does it perform?

Every figure here is NVIDIA's own; we have not tested the model.

  • English. NVIDIA says Transcribe is "first among open-weight models on the Hugging Face Open ASR Leaderboard for English." The blog gives "5.33% average WER." The model card gives a "Mean WER" of 4.71 on "9 public sets," with "Results captured on October 12, 2026." NVIDIA does not explain the gap between the two figures.
  • Multilingual and long-form. The blog gives "4.62% mean WER across seven languages" and "10.98% mean WER" for long-form audio, "with API models excluded."
  • Speaker attribution. The blog says it "posts the lowest overall diarization error rate (DER), 18.26% across 21 languages" on David AI's leaderboard. The card notes "Transcribe's training data includes data licensed from David AI. The DAI-ASR-I18N evaluation set was not part of it."
  • Artificial Analysis. The blog gives "2.8% AA-WER." NVIDIA's two posts describe the rank differently: the blog says it "ties for first among open-weight models," while the NVIDIA blog says "No. 2 among open weight models."
  • Fine-tuning. In a medical case study on recordings from Indian doctors and medical students, NVIDIA says "fine-tuning cuts held-out WER from 22.44% for the best-prompt baseline to 12.17%." This is a transcription benchmark, not a clinical accuracy claim.

#Where can you get it?

The blog says to "download the checkpoint from Hugging Face or ModelScope," self-host with "vLLM or SGLang-Omni," or "use a hosted endpoint from Baseten, DeepInfra, or Simplismart." The card's reference setup runs on "one H100 80 GB."

Local use depends on smaller versions. NVIDIA says "Quantized versions are coming soon, so you can run Transcribe locally on your RTX Spark, DGX Spark, or Jetson AGX Thor." On October 29, 2026, the public Hugging Face collection listed only the BF16 model and its DFlash2 draft model. NVIDIA's throughput claim of "100 hours of audio in under three minutes" on one RTX PRO 6000 Blackwell Server Edition GPU used "the forthcoming NVFP4 quantized version."

#How does it fit the market?

TechCrunch said "These tools target developers building apps for business meetings and medical transcription." It noted that "Mistral shipped Voxtral 3 earlier this month; ElevenLabs launched Realtime for instant speech in voice agents." For another October 2026 voice model, see our Salesforce Launa-Omni story. NVIDIA's last agent launch we covered was its OpenShell Reference Architecture, and Recursive's Flint is another open-weight release with a named license.

Our analysis: One open-weight checkpoint that transcribes, labels speakers, translates and answers questions is useful for anyone building meeting or call tools, and the commercial-use line on the card matters. The benchmark claims need care: NVIDIA's own pages disagree on the English figure and the Artificial Analysis rank. The local-PC story also depends on quantized versions that are not public.

#What is still unclear?

  • Quantized release date. NVIDIA gives no date beyond "coming soon."
  • The English WER gap. No NVIDIA page explains 5.33% versus 4.71%.
  • Hosted pricing. NVIDIA lists no prices for hosted endpoints.