NVIDIA presented Nemotron 3.5 Transcribe on 28 October. Its announcement describes transcription in 40 languages, translation between English and 31 languages, and speaker attribution in 13. It also introduces Granary v2, with data and tools for adaptation. The model targets recorded audio rather than live streaming.

What the model card adds

The official model card lists 34 billion total parameters, including the audio encoder, with around 4.3 billion active. Its reference deployment uses an H100 with 80 GB of memory. The documentation distinguishes tasks and language support, and provides configuration guidance. These details describe a documented setup, not a universal minimum specification.

How we would evaluate it on Linux

Our proposed first trial would use a short recording with a transcript you have checked yourself. Write down separately which words were wrong, whether the correct person received each sentence, and whether the output format suits the intended application. Keeping those questions separate makes it easier to identify what needs improvement.

Then add a more difficult recording resembling your real work. Keep its expected result out of any tuning examples. Compare the same inputs and instructions before changing settings, and record the exact software and model versions. This is an evaluation plan, not a report of tests we have performed.

Decide where the trial belongs

For a small Linux server, our recommendation is to establish a resource budget before downloading a large checkpoint. Consider a separate test machine or a hosted trial if the documented configuration is outside that budget. Compare the effort of operating the system with the value of the task it would perform; running locally is a deployment choice, not an objective in itself.