Nvidia published Nemotron 3.5 Transcribe on Hugging Face on 28 October 2026, an open-weight speech model that transcribes 40 languages, according to its Hugging Face model card. If your product records calls, this is a serious option to run on your own servers. Before it hears a single customer call, decide what you keep, for how long and who can read it. Open weights move the model in-house. They do not decide your retention policy for you.
What is Nemotron 3.5 Transcribe?
Nemotron 3.5 Transcribe is an offline speech model from Nvidia that turns recorded audio into text, with translation and speaker labels in the same model. The model card describes it as "an offline speech and audio model that processes recordings." Offline here means it works on finished recordings, not a live stream.
Under the hood, the model card describes a hybrid of Mamba-2 and Transformer layers using a mixture of experts, with 34 billion parameters in total and 4.3 billion active, counting the audio encoder. Only a small part of the model runs for each step, which keeps the cost of running it lower than its size suggests.
Offline also tells you what it is not for. If you need live captions during a call, look elsewhere. Nvidia's Nemotron Speech collection on Hugging Face lists a separate streaming model, nemotron-3.5-asr-streaming-0.6b, for that job. Pick the model for the moment the text is needed, during the call or after it.
What can it do?
More than plain transcription. These are the figures I could confirm.
- 40 languages for transcription, per the model card.
- Recordings up to one hour. The model card says it was trained with recordings up to an hour long.
- Translation in both directions between English and 31 languages, as The Decoder reports from Nvidia's material. The model card lists the languages and says translation "covers both directions between English and the listed language."
- Speaker labels for up to eight speakers, according to The Decoder, with speaker identification in 13 languages.
Nvidia's Nemotron 3.5 Transcribe collection lists BF16, FP8 and NVFP4 versions, so you can trade some precision for memory if your hardware is tight.
What licence does it use?
The OpenMDW-1.1 licence. The model card says the model "is ready for commercial or non-commercial use," and links the OpenMDW 1.1 licence text. Read it with your lawyer before you ship. A permissive licence is good news. It is not the same thing as a review.
Why does open weights not settle the privacy question?
Because the risky part of speech is not the model. It is the audio.
A customer call contains names, addresses, card numbers read out loud, health details, a child crying in the background. With a hosted API, at least the question "where does this audio go?" is forced on you by a vendor contract. With open weights, it is easy to skip. The model runs on your machine, it feels private, and nobody asks.
That feeling is the trap. Running the model yourself means you are now the one storing the audio, the transcripts, the translations and the speaker labels. Every one of those is personal data. Every one needs a rule.
The card is also clear about its own training data. It says "User interactions with a deployed product or model were not used to train the model." That is a good sign about where the model came from. It tells you nothing about where your customers' audio will go next.
Nvidia's own card points at the second problem too. It says the speaker labels "do not establish real-world identity." A label saying "speaker 2" is not consent from speaker 2. If your product turns those labels into named people, that is your decision and your responsibility, not the model's.
What should the retention decision cover?
- Audio. Do you keep the recording after transcription? If yes, for how long, and why? Deleting it after a successful transcript is a strong default.
- Transcripts. Who can search them? Are they in the same database as everything else, with the same access rules?
- Translations and labels. These are new data. Treat them like the transcript, not like a cache.
- Logs. Inference servers love to log inputs when something fails. Check that debugging does not quietly keep the audio you meant to delete.
- Training. Will you fine-tune on customer calls? If so, say so in your terms and get consent. If not, write that down too.
None of this depends on the model. It is the same list you would write for any speech system. The difference is that with open weights, nobody else will ask you to write it.
Is it worth running yourself?
For many teams, yes, with eyes open. Keeping audio inside your own network is a real advantage for regulated clients. It removes one vendor from your data map. It lets you control cost once volume is high.
The cost is operations. You need GPUs, monitoring, upgrades and someone who understands what happens when a recording is an hour and five minutes long. You also need your own evaluation. Vendor accuracy figures are measured on vendor test sets. Your callers have their own accents, their own background noise and their own jargon. Test on a sample of real, consented calls before you switch. Measure the errors that hurt you, such as wrong names and wrong numbers, not only the overall score.
What would I do this week?
If you already transcribe calls with a hosted service, write down your current retention rules first. Then run Nemotron 3.5 Transcribe on a small consented sample and compare it with what you have. If you do not have retention rules, stop there and write them. That is the more urgent job.
Where Alongside fits
Choosing a speech model, and writing the retention rules before it touches customer calls, is the kind of work covered by Alongside's technical advisory.