News · 2026-10-29

Nvidia releases multilingual Nemotron speech transcription model

Nvidia released Nemotron 3.5 Transcribe on October 28 as an offline speech model supporting 40 languages and recordings up to one hour. The company’s project page and model card establish the base release, while a circulating claim of a much smaller community derivative remains unverified. The practical news is a multilingual transcription model with inspectable release terms, rather than a proven breakthrough in shrinking it.

Key facts

• Nvidia’s release supports 40 languages and recordings up to one hour.
• The model card states about 34 billion total parameters and 4.3 billion active parameters, including the audio encoder.
• Nvidia identifies the base weights as using a 16-bit floating-point format and states OpenMDW-1.1 terms.
• Primary source: Nvidia’s October 28 project page.

The release’s most useful number is language coverage. A transcription service becomes more useful across a multilingual organization when it can process more than the dominant language, but coverage is not the same thing as equal accuracy. A language listed on a model card still needs evaluation against the accents, recording conditions, vocabulary, and speakers that a user actually expects to encounter.

Nvidia’s model card identifies the artifact as “NVIDIA-Nemotron-3.5-Transcribe-30B-A4B.” The branded 30B-A4B shorthand should not override the card’s more detailed total and active counts. Nor should either count be converted into an asserted download size or a claim that the model fits on a particular graphics card.

The verified weight-download size is unavailable in the checked record, and the permitted file-listing double-check did not return accessible data. The dossier also does not establish a minimum or recommended runtime GPU-memory requirement. Those limits are stated here explicitly because disk storage, resident weights, temporary computation, and recording-dependent state are different quantities. There is no defensible exact fit claim to make from the available evidence.

The distinction between total and active parameters is nevertheless important. A selectively activated model can use only part of its learned capacity for a particular computation while retaining a larger collection of available weights. Think of a multilingual reference library: consulting a few volumes for one question does not mean the other shelves have ceased to occupy space. Our mixture-of-experts lesson explains why relatively low active computation can coexist with substantial storage demands.

Offline transcription also needs a plain definition. The model processes recorded speech into text; the documented support for recordings up to an hour does not by itself establish a low-latency, continuously interactive voice assistant. A reader choosing software for live captions or a phone conversation should not infer streaming behavior from a batch transcription release. Our automatic speech-recognition lesson describes the broader job of turning an audio signal into a sequence of words.

The dramatic claim circulating around the release concerned pruning: a community headline described a reduction from 67.5 GB to 10 GB without retraining, often rounded to a sevenfold reduction. The dossier did not locate an authoritative derivative artifact or write-up establishing those sizes, the method, the author, the derivative’s license, or its quality. Those headline sizes are therefore unverified claims, not measurements of the base model’s official files.

That distinction is more than an editorial technicality. A smaller file might remove experts, change numerical precision, alter the packaging, or omit components. Each possibility changes what must be tested. A result that preserves performance on a small English sample might still harm other languages or difficult acoustic conditions. Without a derivative artifact and a clear evaluation protocol, there is no reliable way to tell what was preserved.

The base model’s release terms also should not be inherited by an unidentified derivative without checking. Nvidia states OpenMDW-1.1 terms for the base release. That establishes the company’s declared terms, rather than a blanket assertion that every redistributed package has identical rights, contents, and obligations. The model card is the appropriate starting point for evaluating the actual artifact.

For builders, the useful next step is to test the official release against representative recordings and inspect its documented execution path. The significant development is that Nvidia has shipped a multilingual offline transcription model with a public project page and model card. Its limitations in this briefing are equally concrete: runtime memory is unstated in the checked evidence, download size could not be verified, and the viral pruning claim has not met the verification bar.

Primary source, verified: read the paper →

Ground Truth. (2026, October 29). Nvidia releases multilingual Nemotron speech transcription model. Ground Truth. https://groundtruth.day/news/nvidia-nemotron-3-5-transcribe-release.html

@misc{groundtruth:nvidia-nemotron-3-5-transcribe-release,
  title  = {Nvidia releases multilingual Nemotron speech transcription model},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/nvidia-nemotron-3-5-transcribe-release.html}
}

Topics: nvidia · speech-recognition · open-weight-models · model-releases · local-ai