Nvidia Releases Open Nemotron 3.5 Transcribe Model

On October 28, 2026, Nvidia released Nemotron 3.5 Transcribe, an open-weight model that transcribes recorded audio in 40 languages. It can be used commercially, but it needs a powerful GPU. It matters to anyone who publishes videos, podcasts or recordings and wants to turn them into text.

What we know

What it is: according to Nvidia's official model card on Hugging Face, an open-weight speech and audio model that works offline on recordings. It is not for live audio: the card says it is not a real-time transcription service.
Languages: 40 transcription languages. Nvidia puts English, Spanish, French, German, Portuguese and several others in its top quality tier. It also translates to and from English for most of them.
Output: text, with optional word-level timestamps and speaker separation (labels such as "speaker 0" that do not identify the real person). It accepts recordings of up to one hour in WAV, FLAC, MP3 or OGG.
Size: about 34 billion parameters in total, with about 4.3 billion active at each step, per the card. TechCrunch reports 30 billion.
Hardware: the card targets a single 80 GB H100 GPU and lists other compatible hardware, from datacenter cards to the RTX 5090 (only in its most compressed format) and DGX Spark systems.
License: OpenMDW-1.1. The card says the model is ready for commercial and non-commercial use.
Performance: according to Nvidia, a 4.71% word error rate on the mean of the English Open ASR Leaderboard, versus 4.94% for the next open model it cites (Qwen3-ASR-1.7B). That is Nvidia's own measurement on public benchmarks, not an independent test, and it says nothing about poor-quality audio or specific accents.

What changes and what doesn't

What changes is that there is now an open model, cleared for commercial use, that separates speakers and returns timestamps, with strong results claimed across major languages. For a developer or a service provider, that makes it possible to transcribe large volumes on their own servers instead of paying per minute to an outside service.

Nothing changes for a site on shared hosting: the model needs a GPU that kind of server doesn't have. It doesn't replace live transcription, which the card rules out. And a low error rate on benchmarks doesn't guarantee publishable text: names, brands and technical terms still need a human check.

How to tell if it affects you

If you publish video or podcasts, count the hours. For a few hours a month, a paid transcription service or your video platform's automatic captions are simpler. Running your own model only pays off at high volume, or with audio you don't want to send outside your servers.
Put the text on the page, not just in the video. Search engines read the text on your site far more easily than the audio in a video. A reviewed transcript or a summary under the video helps the page be understood and found, and improves accessibility.
Review before publishing. Fix names, numbers and industry terms. An unreviewed automatic transcript can put mistakes on your site under your name.
If a provider offers it, ask three questions. Where the audio is processed, whether it is kept afterwards, and whether the model is this one or another. The license allows commercial use, but what happens to your recordings depends on who runs it.
If you have a technical team, test it on your own audio. Ten minutes of a real recording, with noise, accents and proper names, tells you more than any results table.

Sources

Updates: if independent tests or a hosted API with pricing appear, they will be added here with a link.