NVIDIA Nemotron 3.5 Transcribe

What it is

Nemotron 3.5 Transcribe is a speech transcription model released by Nvidia on 28 October 2026. It is free to use and its weights are open, meaning others can download and run the model themselves rather than only reaching it through a paid service^[1].

It belongs to the same family of open voice tools as Parakeet, Nvidia's earlier transcription model, which has 1.1 billion parameters^[1]. Nemotron 3.5 Transcribe is considerably larger, at 30 billion parameters^[1].

What it does

The model turns speech into text in up to 40 languages^[1]. It also translates: 31 languages can be translated to and from English^[1]. Transcription and translation therefore sit in the same model rather than in two separate tools.

It is built for recorded audio rather than live streams, an important distinction for anyone planning to use it. Nvidia says the model works with recordings, not livestreams^[1]. Within that setting it handles long material, since a single recording can run up to an hour, and it can tell apart as many as eight speakers in the same audio^[1]. That speaker separation matters for meeting recordings, interviews and panel discussions, where the useful output is not just a wall of words but an indication of who said what.

Built to run on one chip

A 30-billion-parameter model is normally something you run on a server with several accelerators. Nvidia's design goal here was different: the model is meant to run on a single chip^[1]. The range of hardware Nvidia names stretches from the Jetson AGX Thor robotics controller and the DGX Spark desktop machine to Blackwell data centre chips^[1]. In other words, the same model is intended to work on a small device next to a robot as well as in a data centre.

That choice shapes who can use it. A model that fits on one chip can be run locally, without sending audio to an outside service, which matters for applications where the recording is sensitive or where a network connection is unreliable.

Data and benchmarks

Alongside the model, Nvidia released version 2 of Granary, its open-source speech dataset^[1]. Speech models need large amounts of transcribed audio to train on, and an openly released dataset lets outside researchers train and test their own systems on the same material.

Nvidia claims the model beats its rivals on benchmarks, including Hugging Face's Open ASR Leaderboard^[1]. That leaderboard is a public comparison of automatic speech recognition systems, and a claim of leading performance there is a claim about accuracy relative to other openly compared models. Such claims are Nvidia's own; independent users can check them against the leaderboard.

Where it fits

Open-weight speech models have become a busy area, and Nvidia is positioning this one as the larger, more capable member of a family that already includes Parakeet^[1]. The combination is characteristic of how open models are usually released: weights that anyone can download, a dataset that supports further work, and a claim of benchmark leadership to make the case for switching.

What the release does not cover is live audio. Nvidia describes the model as working with recorded audio rather than livestreams^[1], so real-time transcription, such as call centres, live captioning and voice assistants, is outside the scope it was announced for. Users with an hour-long recorded meeting in one of 40 languages, or a recording in one of 31 languages that needs to reach English, are the audience the release is aimed at.