NVIDIA released Nemotron 3.5 Transcribe on October 28. It is an open-weight speech-to-text model: anyone can download it from Hugging Face or ModelScope and run it on their own hardware. The model card says it is ready for commercial or non-commercial use under the OpenMDW-1.1 license, so there is no fee for the model itself.

What it does

Transcription in 40 languages, with timestamps for every word.
Speaker labels for up to eight speakers, in 13 of those languages.
Translation of speech in 31 languages to or from English.
Prompt control: you can steer the output with plain-language instructions and ask questions about a recording (this part was trained in English).

Two limits matter. Recordings can be up to an hour long, and the model only works on recorded audio. NVIDIA's blog says it is built for recorded audio, not live streaming, and the model card calls it an offline model.

Numbers and hardware

The model has about 30 billion parameters, of which roughly 4 billion are active for each token. NVIDIA reports an average word error rate of 5.33% on the English track of Hugging Face's Open ASR leaderboard, first among open-weight models, and says a single RTX PRO 6000 Blackwell server GPU can transcribe 100 hours of audio in under three minutes in an optimised setup. These are NVIDIA's own figures.

This is not a laptop model. The model card's reference deployment is one H100 GPU with 80 GB of memory; quantised versions are meant for smaller NVIDIA systems such as DGX Spark and Jetson AGX Thor. Without such hardware, you can use hosted versions from Baseten, DeepInfra and Simplismart. Alongside the model, NVIDIA published Granary v2, an open speech dataset of about 463,000 hours in 48 languages.

Who it matters to

Developers who build transcription into their own products, and organisations that cannot send recordings to an outside service, get a strong free option they can run in-house. Fine-tuning helps in specialised fields: in NVIDIA's medical example, the word error rate fell from 30.8% to 11.8% after fine-tuning on about 245 hours of audio.

What this means if you need transcription

If you are a developer choosing an open model, Nemotron 3.5 Transcribe is now a serious alternative to OpenAI's Whisper, especially if you want speaker labels and translation from the same model. Test it on your own audio before switching: leaderboard scores may not reflect your recordings.

If you just want notes from your meetings, this is not an app. Ready-made tools such as Otter.ai and Granola record the call and write the notes for you with no setup, and they work while the meeting is happening.