Nvidia Nemotron 3.5 Transcribe: Free Open AI Model Turns Speech Into Text in 40 Languages
If you’ve ever spent a Sunday typing up a one-hour interview, this one’s for you.
Nvidia, the company whose chips power most of the AI boom, has released Nemotron 3.5 Transcribe. It’s a speech-to-text model that anyone can download. According to its Hugging Face page, it was released on 28 October 2026. And it does a lot more than turn audio into words.
What does it actually do?
In Nvidia’s own words, it’s a “speech-augmented language model”, basically a language model that can listen. One model handles all of this, according to Nvidia’s research page and its Hugging Face blog:
- Transcribes 40 languages, either word for word or as a cleaner written version.
- Translates 31 of them to or from English.
- Works out who said what in 13 languages, for up to eight speakers.
- Handles recordings up to one hour long.
- Adds word-level timestamps, detects the language and can answer questions about the audio.
You tell it what you want in plain English. A clean transcript, a verbatim one or a translation all come from the same model, with no retraining.
The feature we like most
Names.
Auto-captions are terrible with names. Brand names, product codes and the name of the guest you spent a week booking all come out wrong. Nemotron lets you give it a list of the names and terms you expect to hear. Nvidia says this cut named-entity errors six-fold on a benchmark called ContextASR-Bench.
That might sound small. It isn’t. Having to fix every name is what makes most transcripts unusable without a full re-edit.
Is it really that good?
Nvidia says it ranks first among open-weight models on Hugging Face’s Open ASR Leaderboard for English, and it also claims top results in speaker-labelling comparisons from David AI and VoiceArena. Those are strong numbers, but they’re benchmarks Nvidia chose to report. Noisy audio, heavy accents and three people talking at once are another matter, so test it on your own recordings.
The catch
There are two.
First, it doesn’t do live audio. Nvidia says clearly that it’s an offline model for recordings, so it’s for podcasts, meetings and interviews, not live captions on a call.
Second, you need serious hardware. The model has about 34 billion parameters, though only around 4.3 billion are active at a time thanks to a “mixture of experts” design. The full version is meant for a single H100 80 GB data-centre GPU, and the lighter FP8 and NVFP4 versions target Nvidia’s RTX PRO 6000 Blackwell cards. Nvidia also mentions devices like DGX Spark and Jetson AGX Thor for running it locally. In other words, this won’t run on your laptop.
What it means if you’re in India
There’s no regional lock. The weights are on Hugging Face (Nvidia also lists ModelScope), and the licence, OpenMDW-1.1, allows commercial use according to the model card. Any Indian startup, newsroom or agency can pick it up today.
If you just want to try it without any hardware, Nvidia’s blog points to a demo on Hugging Face. Upload a short clip and see how it does. Before you plan anything around Hindi or another Indian language, check the language support list on the model card, since coverage varies by task.
The bigger deal is for companies with sensitive audio, like call centres, clinics and legal teams. Because it’s open, they can run it on their own servers and keep recordings in-house instead of sending them to a paid cloud service.
Our take
Nvidia isn’t giving this away out of kindness. Every developer who builds on its open models is likely to end up buying its GPUs, and that’s the business plan. Still, users come out ahead. You probably won’t open “Nemotron” yourself, but expect it in the apps you use: sharper captions, better meeting notes and fewer mangled names. If you make podcasts or videos, try the demo this week with your messiest recording.
FAQ
Is Nvidia Nemotron 3.5 Transcribe free?
The weights are openly available on Hugging Face under the OpenMDW-1.1 licence, and the model card says it’s fine for commercial and non-commercial use. You still need capable hardware to run it yourself.
How many languages does it support?
Nvidia says it transcribes 40 languages, translates 31 to or from English and labels speakers in 13.
Can it transcribe live calls?
No. It’s an offline model built for recorded audio.
Sources: Nvidia Research, Nvidia’s Hugging Face model card and blog, and AI Early Edition.
