NVIDIA releases Nemotron 3.5 Transcribe and Granary 2

Background

Transcription, turning speech into text, is one of the more settled uses of AI in the workplace. When a team picks a speech model, three questions tend to come up: which languages it covers, whether the team can run it on its own machines, and how well it copes with the vocabulary of a particular profession. Nvidia's release speaks to all three.

What was released

On 28 October 2026 Nvidia released Nemotron 3.5 Transcribe, together with the second version of its Granary speech dataset ^[1]. The model is open-weight and free, and covers up to 40 languages ^[1]. Open-weight means the trained parameters are published so that a user can download and run the model directly, rather than only reaching it through someone else's service; free describes the terms of the release.

Who it is aimed at

Nvidia targets business meeting and medical transcription apps, two established workplace uses of AI ^[1]. Meeting transcription turns a discussion into a text record that can be searched and shared; medical transcription turns spoken notes into written records. In both, a mistake in a name, a number or a term carries weight, which is why vocabulary is the interesting part of the release.

Key details

The model has 30 billion parameters, and Nvidia pitches it as running on a single chip across Jetson, DGX Spark and Blackwell hardware ^[1]. Parameters are the internal numbers a neural network adjusts during training; the count is a rough measure of capacity, and it also sets how much memory and compute a deployment needs. Running on a single chip rather than a cluster of servers changes where a transcription system can sit.

Nvidia says developers can combine the model, the Granary dataset and its NeMo Speech tools to recognise industry jargon in each language ^[1]. Jargon is the vocabulary a profession uses internally — internal product names, department abbreviations, drug names — the layer where a general-purpose model has the least to go on. A dataset and a toolchain are what let a team adapt a model to that layer.

A speech dataset is used in two ways: models learn the mapping between audio and text from paired recordings and transcripts, and held-out recordings measure how well a model does on speech it has not seen. Shipping a dataset version together with a model and a toolchain points at adapting the model to the vocabulary of a particular language and industry.

A busy month

The release landed in a busy month for voice models, alongside Mistral's Voxtral 3, ElevenLabs Realtime and Salesforce's Launa-Omni ^[1].

What the release amounts to

Taken together, the terms are: free, open-weight, up to 40 languages, single-chip deployment, plus a dataset and a toolchain shipped alongside ^[1]. For a team that wants to build transcription in-house rather than call a service, those are the conditions that decide whether a model is something they can download, adjust and run themselves.