Nvidia’s Nemotron 3.5 model lets companies prompt and adapt transcripts

Nvidia has released Nemotron 3.5 Transcribe and Granary v2, a pair of open models for intelligent speech-to-text in dozens of languages. The key feature is that these models can be prompted and adapted, letting companies fine-tune transcription outputs for their own workflows.

Why it matters

Speech-to-text has been a sticking point for businesses dealing with international teams, call centers, or content that has to move between languages. Most cloud transcription has meant paying a vendor, accepting black-box models, and settling for generic outputs. With Nemotron 3.5 Transcribe and Granary v2, Nvidia says companies can now run multilingual transcription on their own infrastructure and prompt the model for specific needs, like emphasizing domain vocabulary, formatting, or even summarizing.

This is not the first open multilingual transcription model, but the addition of promptability and adaptation is the claim to watch. According to Hugging Face, these models are built to be flexible, rather than forcing everyone to use the same fixed pipeline. For industries with compliance rules or specialized lingo, that means more direct control over how voice data becomes text. For businesses that have hesitated to move sensitive audio offsite, on-premise options are now practical.

Promptable transcription: what Nvidia released

Nemotron 3.5 Transcribe is Nvidia’s new large language model for converting speech to text, supporting dozens of languages. The model is built to be prompted, meaning users can specify instructions or preferences at runtime. For example, a contact center could prompt the model to prioritize customer names, or a legal firm could ask for transcripts in a particular format. Granary v2 is a companion dataset for training or evaluating transcription models, which Nvidia says is larger and more diverse than previous public sets.

Companies can deploy Nemotron 3.5 Transcribe on their own hardware or in the cloud, according to Nvidia. This avoids the need to send sensitive audio to external providers. The company also claims that, because the model is open, it can be adapted for specific domains or integrated into custom applications. Hugging Face reports that the model’s multilingual support covers dozens of languages and dialects, though live audio streaming is not in scope for this release. For a look at how the field is shifting, see our coverage of Nvidia's previous free model.

What this means if you run a business

If your company records meetings, runs a support line, or produces content in multiple languages, Nemotron 3.5 Transcribe means you can keep voice data in-house and tailor transcription to your needs. This could cut costs for firms paying per-minute fees to cloud services, and it puts more control in IT and compliance hands. For regulated industries, like healthcare, finance, or law, the ability to adapt transcription workflows and vocabulary is a direct answer to recurring headaches about accuracy and confidentiality.

Expect a learning curve: promptable models reward those who invest time in tuning instructions and adapting outputs. But for companies with the technical resources, the trade-off is clear: more control, less vendor lock-in, and a path to workflows that actually match your business rather than the other way around.