Nemotron 3.5 ASR

NVIDIA's streaming speech-to-text model: 40 language-locales in one 0.6B model, with the delay set per request from 80 ms to 1.12 s.

In short. Nemotron 3.5 ASR turns live speech into text as it is spoken. One small model transcribes 32 language-locales out of the box and knows 8 more once you fine-tune it. You choose at run time how long it waits before committing words, trading a little accuracy for a shorter delay.

Summary

Nemotron 3.5 ASR is the multilingual version of NVIDIA's English streaming recognizer, and the card gives its Hugging Face release date as 4 June 2026. It has about 638 million parameters, a 24-layer cache-aware FastConformer encoder and an RNN-T decoder. Cache-aware means the encoder keeps the results of earlier audio chunks and processes only the new audio, instead of re-reading overlapping windows the way buffered streaming does. A language-ID prompt tells the model which language to expect, or it can detect the language itself and tag each utterance with it. The chunk size, and so the delay, is picked at inference time from 80, 160, 320, 560 or 1,120 milliseconds, with no retraining. Output has punctuation and capitalization.

The card counts 40 language-locales in three tiers. Nineteen are transcription-ready, including English (US and UK), Spanish, French, Italian, Portuguese, Dutch, German, Turkish, Russian, Arabic, Hindi, Japanese, Korean, Vietnamese and Ukrainian. Thirteen are broad-coverage, including Polish, Swedish, Czech, Mandarin and Finnish. Eight are adaptation-ready, which means the tokenizer knows them but you must fine-tune on your own data before they transcribe usefully: Greek, Lithuanian, Latvian, Maltese, Slovenian, Hebrew, Thai and Norwegian Nynorsk. For English only, NVIDIA points to its separate English streaming model, which is released under the NVIDIA Open Model License.

NVIDIA reports word error rates on FLEURS, and character error rates for Japanese, Korean and Mandarin. Across the 19 transcription-ready locales the average is 10.38% at the 80 ms setting and 8.84% at 1.12 s when you name the language, and 11.14% at 80 ms with automatic language detection. English scores 9.43% at 80 ms and 7.91% at 1.12 s, and Spanish 4.87% and 4.11%. NVIDIA warns that its text normalization is imperfect across locales, so real quality may be somewhat better than these figures. On throughput, it says one H100 serves about 17 times more concurrent streams than its Parakeet RNNT 1.1B multilingual model at the 80 ms setting (240 against 14) and 6 times more at 1,120 ms (2,400 against 400). These are NVIDIA's own measurements.

Training used between 10,000 and 1 million hours of audio across the 40 locales, from NVIDIA's own multilingual set plus Granary, Multilingual LibriSpeech, Common Voice, FLEURS and VoxPopuli. Some labels came from an ensemble of other recognizers, and punctuation was added with Qwen3-32B. The license is OpenMDW-1.1. The model runs in NVIDIA NeMo, in Hugging Face Transformers from version 5.13.0, and through the lightweight NeMo-Speech.cpp runtime, which can use an 8-bit GGUF file of 0.74 GB from the same repository.

Good for

  • Voice agents and live captions that need words within a fraction of a second
  • Serving many simultaneous audio streams from one GPU
  • Mixed-language traffic, since the model can detect and tag the language of each utterance
  • Fine-tuning for lower-resource languages, because eight more locales are already in the tokenizer

Watch out for

  • Eight of the 40 locales need fine-tuning on your own data before they transcribe usefully
  • For English only, NVIDIA recommends a separate English model that uses a different license
  • Accuracy varies a lot by language, from under 5% in Spanish to about 16% in Ukrainian at the 80 ms setting
  • Automatic language detection is less accurate than naming the language, and the gap is large for Hindi, Russian and Ukrainian
  • The error rates and throughput figures are NVIDIA's own

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Nemotron 3.5 ASR. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
Nemotron 3.5 ASR Streaming 0.6B FP32 638M 2.55 GB 2.93 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.

More about OpenMDW-1.1

See also

Frequently asked questions

What is Nemotron 3.5 ASR?
Nemotron 3.5 ASR turns live speech into text as it is spoken. One small model transcribes 32 language-locales out of the box and knows 8 more once you fine-tune it. You choose at run time how long it waits before committing words, trading a little accuracy for a shorter delay.
Can I use Nemotron 3.5 ASR commercially?
Yes. Nemotron 3.5 ASR is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Nemotron 3.5 ASR open source?
Nemotron 3.5 ASR is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
Can I run Nemotron 3.5 ASR on my own computer?
Yes. The smallest variant, Nemotron 3.5 ASR Streaming 0.6B, needs about 2.93 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Nemotron 3.5 ASR released?
Nemotron 3.5 ASR was released by NVIDIA on June 4, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Parakeet TDT 0.6B v3 NVIDIA 627M – CC BY 4.0 (Permissive) 2.89 GB Laptop or phone
Nemotron 3.5 Transcribe NVIDIA 34B 4.3B active 64K OpenMDW-1.1 (Permissive) 25 GB Gaming GPU
Voxtral Realtime Mistral AI 4.43B 128K Apache 2.0 (Permissive) 10 GB Laptop or phone
Granite Speech 5.0 IBM 470M – 2.31B – Apache 2.0 (Permissive) 1.09 GB Laptop or phone