Nvidia released Nemotron 3.5 Transcribe on October 28, 2026. It’s a free, open-weight model that turns recorded audio into text in 40 languages and tells up to eight speakers apart. It’s built for batch jobs like meeting recordings, not live captions, and it needs a big GPU to run.
Here are the facts worth knowing before you click download.
- Released October 28, 2026, with weights on Hugging Face, per TechCrunch.
- Transcribes 40 languages and translates 31 of them to or from English.
- Handles recordings up to one hour and labels up to eight speakers.
- Recorded audio only. There is no live or streaming mode.
- Free to download, but the weights are too large for most home PCs.
The short version: it’s great news for developers and companies, and mostly a “wait for smaller versions” story for everyone else.
What Nemotron 3.5 Transcribe actually does
Most speech-to-text tools do one thing: write down the words. This one bundles several jobs into a single pass. It transcribes, translates, separates speakers, and can answer questions about what was said in the recording.
You can also hand it a short list of hint words, such as drug names or product terms. That helps it spell specialist vocabulary correctly instead of guessing. According to The Decoder, speaker labels work in 13 languages, and the question-answering feature was trained mostly on English.
Nvidia pitches it at meeting-notes apps, call-center review, interview archives and medical dictation. Think of a one-hour team call: you’d get a written transcript, a rough speaker for each line, and an English translation if people spoke another language.
Feature support varies by language, so a language that transcribes well may not get every extra. Check your language before you plan a workflow around it.
Can your computer run Nemotron 3.5 Transcribe?
Probably not, and that’s the catch. Parameter counts differ slightly between outlets (roughly 30 to 34 billion, with about 4 billion active at a time), but the weights still have to fit in GPU memory. Tech Times lists these approximate sizes.
| Format | Approx. memory needed | Who can run it |
|---|---|---|
| BF16 (full precision) | About 68 GB | Data-center GPUs |
| FP8 | About 34 GB | High-end workstation cards |
| NVFP4 (smallest) | About 17 GB | Large pro cards, DGX Spark, Jetson AGX Thor |
Per the same report, a typical 8 to 16 GB consumer graphics card can’t load it in any format. If you want a small model that fits on a phone, look at Liquid AI’s LFM3 on-device models instead.
Speed is the upside. Nvidia says one RTX PRO 6000 Blackwell card can process 100 hours of audio in under three minutes at the smallest format. That figure comes from Nvidia, so treat it as a best case.
How accurate is it?
Nvidia says the model ranks first among open-weight models on Hugging Face’s English Open ASR Leaderboard. It’s also said to lead a test built around overlapping voices, where people talk over each other.
Reports disagree on the exact leaderboard average. TechCrunch gave no score, while The Decoder and Tech Times quote slightly different figures. Most numbers trace back to Nvidia, and Tech Times noted they weren’t independently verified at publication.
For a similar reason, we’d treat any “beats the rivals” headline carefully, as we did with the Nanbeige4.3-33B benchmarks. Run a few of your own recordings before you trust it with anything important.
It also lands in a busy stretch for voice AI. TechCrunch points to recent releases from Mistral (Voxtral 3), ElevenLabs (Realtime) and Salesforce (Launa-Omni). Many of those aim at live conversation, while Nvidia’s pick is big, accurate batch transcription.
Free to use, with some fine print
The weights are open and free. Per Tech Times, they ship under the OpenMDW-1.1 license, which allows training, modification, redistribution and commercial use. Whether it counts as “open source” under the Open Source Initiative’s definition is still under review.
One interesting angle is privacy. Running the model on your own hardware means audio never leaves your servers. The same report says this can avoid some cloud-vendor agreements in healthcare, but it advises talking to legal counsel before relying on that.
Nvidia also showed that fine-tuning helps. In a medical test, error dropped from 23.8% to 11.8% after about 14 GPU hours, though general English got slightly worse. Nvidia recommends spot-checking names, speaker changes and answers.
Frequently Asked Questions
Is Nemotron 3.5 Transcribe free?
Yes. The model weights are free to download from Hugging Face. You still pay for the hardware or cloud GPU time to run it.
Can it transcribe live calls or streams?
No. It’s built for recorded audio only. For live captions you’d need a different, real-time model.
Which languages does it support?
It transcribes 40 languages and translates 31 of them to or from English. Speaker labels reportedly work in 13 languages.
Can I run it on a regular laptop?
Almost certainly not. Even the smallest format needs about 17 GB of GPU memory, according to Tech Times. Try the Hugging Face demo instead.
Our take
If you build transcription tools or handle lots of recorded meetings, Nemotron 3.5 Transcribe is worth a weekend test, especially for private audio you don’t want in the cloud. If you just want to transcribe a podcast on your laptop, skip it for now and wait for smaller models. Either way, test it on your own recordings first, because benchmark wins and real-world accuracy aren’t the same thing.
Leave a Reply