NVIDIA Releases Nemotron 3.5 Transcribe and Granary v2 on Hugging Face
NVIDIA has published a Hugging Face blog post introducing Nemotron 3.5 Transcribe and Granary v2, a multilingual transcription release it describes as something you can prompt and adapt. The post links two models, two datasets and two Spaces.
- NVIDIA published a Hugging Face post on Nemotron 3.5 Transcribe and Granary v2.
- The post is described as covering multilingual transcription you can prompt and adapt.
- It links 2 models, 2 datasets and 2 Spaces, including the Open ASR Leaderboard and a Nemotron Transcribe demo.
- One listed model falls under Automatic Speech Recognition and another under Voice Activity Detection.
- Performance numbers and language coverage are not given in the available text.
What NVIDIA has published
The post is titled "NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt". The title points to two parts of the release: Nemotron 3.5 Transcribe, a speech recognition offering, and Granary v2, which appears to be the data side of the project. The headline also says the transcription can be prompted and adapted, which suggests users can steer the system rather than only run it as a fixed tool.
The text available for this article is mostly the page's navigation and metadata, not the body of the post. It does not describe how prompting works, which languages are covered, or how large the datasets are. Those details are not confirmed here and should be checked in the full post.
Models, datasets and demos
According to the page, the article references 2 models, 2 datasets and 2 Spaces. One model is listed under Automatic Speech Recognition, and the page notes it was updated about one hour before the page was captured. A second listing falls under Voice Activity Detection, a task that detects where speech occurs in audio, and it shows an update date of Sep 24.
The two Spaces are the Open ASR Leaderboard and a demo called Nemotron Transcribe. The leaderboard is described on the page as a place to "Explore speech-to-text model performance across datasets", and the page shows it with 1.51k likes. The Nemotron Transcribe Space is labeled "NVIDIA Nemotron Transcribe Demo", so readers can try the system directly in the browser.
Other NVIDIA posts on the page
The page also lists more posts from the same author. One is titled "Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required" and is dated October 21, 2026. The other is "Train in Simulation, Deploy on a Real Robot: Microduck with Isaac Lab 3.0", also dated October 21, 2026. They show that NVIDIA is using Hugging Face as a channel for a range of technical releases, not only speech work.
What it means for developers
For developers working on transcription, the release combines models, data and live demos in one place. The inclusion of the Open ASR Leaderboard Space gives a route to compare speech-to-text models across datasets, which is useful when deciding whether a new model fits a given use case. The promise of prompting and adaptation, if the full post bears it out, would matter for teams that need transcription tuned to specific domains or languages.
Performance figures, licensing terms and supported languages are not included in the material reviewed, so any assessment of how Nemotron 3.5 Transcribe compares with other systems has to wait for the full blog post and the model pages.
