THE ESSENTIALS IN 30 SECONDS

  • Nemotron 3.5 Transcribe, released October 28, 2026, transcribes 40 languages and handles recordings of up to one hour.
  • Granary v2 covers 463,062 hours in 48 languages as manifests; the audio itself must be fetched from the original corpora.
  • The model works on recorded audio only, and its leaderboard results are Nvidia's own claims.

Nvidia has added a large transcription model to its open Nemotron family. Nemotron 3.5 Transcribe, published on October 28, 2026, is a promptable speech model whose weights are on Hugging Face under the name NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. It arrives alongside Granary v2, a second version of the company's open speech dataset. It should not be confused with Nemotron 3.5 ASR, a small streaming model Nvidia released in June.

What the model does

Transcribe pairs a speech encoder with a hybrid Mamba–Transformer mixture-of-experts language model. Nvidia's Hugging Face post puts it at about 30 billion parameters, of which roughly 4 billion are active per token; some outlets count 34 billion once the audio components are included. It transcribes 40 languages, translates 31 into or out of English, and can attribute speech to up to eight speakers, in 13 languages for that task. Recordings can last up to an hour. Users can steer the output with plain-language instructions, for example asking for a verbatim transcript or listing product names to expect, and the model can answer questions about a recording in English.

According to Nvidia, it fits on a single GPU, from a Jetson AGX Thor board or DGX Spark up to Blackwell data-center chips. Unite.AI reports that the weights use the OpenMDW-1.1 license, which allows commercial use. The company says it ranks first among open-weight models on Hugging Face's Open ASR Leaderboard for English, with an average word error rate of 5.33%.

Granary v2

The dataset's public manifest covers 463,062 hours of speech across 48 languages, drawn from 16 source corpora; the first version covered 25 languages. Granary v2 does not ship audio. It provides transcripts in three forms, task labels and quality metadata, and users download the recordings from each original corpus under that corpus's license. Nvidia stresses that these figures describe the release, not the full training mix of the model.

Caveats

Transcribe is built for recorded files, not live streams. Benchmark claims come from Nvidia, and Unite.AI notes that the leaderboard ranking leaves out API-only services. A medical case study shows the trade-offs of adaptation: fine-tuning on doctors' speech cut errors on that domain by 45.8%, while general English accuracy slipped slightly. Nvidia itself advises checking names and speaker changes against the original audio.