Whisper large-v3-turbo vs Nemotron 3.5 Transcribe: which to use?
Whisper large-v3-turbo is small and widely supported; Nemotron 3.5 Transcribe adds speaker labels and prompt-controlled style. Compare memory, languages and licenses.
Verdict. Whisper large-v3-turbo is the small, well-supported default: MIT license, about 6 GB of GPU memory and many community apps and runtimes. Nemotron 3.5 Transcribe is newer (October 2026) and much larger (34B parameters, 4.3B active, about 22 GB even in its NVFP4 build), but you can ask it in plain language for speaker labels, verbatim or cleaned-up text, and translation. It works only on finished recordings, not live audio. Both developers' accuracy numbers are their own, so test your own audio.
Specifications
| Specifications | Whisper large-v3-turbo | Nemotron 3.5 Transcribe |
|---|---|---|
| Developer | OpenAI | NVIDIA |
| Released | October 1, 2024 | October 28, 2026 |
| Total parameters | 1.54B (large-v3) | 34B (BF16) |
| Active parameters | All | 4.3B |
| Architecture | Other | Mixture of experts |
| Context window | – | 64K tokens |
| Input | Audio | Audio, Text |
| License | MIT (Permissive) | OpenMDW-1.1 (Permissive) |
| Memory for the smallest variant (4-bit) | 6 GB (large-v3-turbo) | 25 GB (NVFP4) |
Performance claims come from the developer and have not been independently verified.
Pick Whisper large-v3-turbo if
- You want the lightest option: about 6 GB of GPU memory for the turbo model, and it runs in many community apps.
- You need sentence-level or word-level timestamps for subtitles.
- You want the MIT license and the widest tool support.
- You need translation into English: use the large-v3 checkpoint, because the turbo model does not translate.
Pick Nemotron 3.5 Transcribe if
- You need speaker labels (available in 13 languages) or a transcript style you can steer with a prompt.
- You work on medical, legal or technical recordings and can supply the terms to expect.
- You need translation, which covers 31 of its 40 languages.
- You have a large GPU (the model card targets a single 80 GB H100, and the NVFP4 build is about 22 GB) and only process finished recordings of up to an hour.