Nemotron 3.5 Transcribe
NVIDIA's open speech-to-text model: 40 languages, speaker labels, translation and prompt-controlled transcript style.
In short. Nemotron 3.5 Transcribe turns recorded speech into text. Unlike older speech-recognition tools, you can tell it in plain language what style of transcript you want: word for word or cleaned up, with speaker names, or translated. It works on finished recordings up to an hour long, not on live streams.
Summary
NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026 under the OpenMDW-1.1 license, together with a fine-tuning toolkit in NeMo Speech and version 2 of its Granary speech dataset.
It is a speech-augmented language model: a Transformer speech encoder and a diarization (who-spoke-when) model feed a mixture-of-experts language model. The card lists about 34 billion parameters in total and 4.3 billion active, including the audio encoder. One model handles transcription, speech translation, speaker-attributed transcription, language identification and audio question answering, and it can return word-level timestamps. Support covers 40 languages, with translation to or from English in 31 of them and speaker attribution in 13, on recordings up to one hour with up to eight speakers.
The transcript style is set by a text instruction, for example verbatim versus written-form numbers, or a list of expected names and terms. This model processes recordings after the fact; it is not a live-streaming recognizer. NVIDIA reports a mean word error rate of 4.71% on the English Open ASR Leaderboard and 4.62% on the multilingual one; these are the developer's own measurements as published on its research page.
The BF16 weights are about 68 GB, FP8 about 35 GB and NVFP4 about 22 GB. The card lists a single H100 80 GB as the deployment target, and GGUF files are available for NVIDIA's NeMo-Speech.cpp.
Good for
- Meeting, call and interview transcripts that need speaker labels
- Multilingual transcription and translation to or from English
- Medical, legal or technical transcripts where you can supply the expected terms
- Teams that want to fine-tune a speech model on their own audio with an open toolkit
Watch out for
- Offline only, since it processes finished recordings, not live audio streams
- It is large for a speech model, with 68 GB of BF16 weights; the NVFP4 build is about 22 GB
- Translation covers 31 of the 40 languages and speaker attribution only 13
- Error rates are NVIDIA's own results; test on your own audio before relying on them
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 Transcribe. 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 |
|---|---|---|---|---|
| BF16 BF16 | 34B 4.3B active | 68 GB | 78 GB | Workstation |
| FP8 FP8 | 34B 4.3B active | 35 GB | 40 GB | Workstation |
| NVFP4 NVFP4 | 34B 4.3B active | 22 GB | 25 GB | Gaming GPU |
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.
See also
- Whisper large-v3-turbo vs Nemotron 3.5 Transcribe: which to use?
- Best open-source speech models for transcription and text-to-speech
- Best open-source speech recognition (speech-to-text) models in 2026
- Best open-source translation models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
- Giant open models now ship in 4-bit: what it means for memory planning
- NVIDIA releases Nemotron 3.5 Transcribe, a promptable speech model
Frequently asked questions
- What is Nemotron 3.5 Transcribe?
- Nemotron 3.5 Transcribe turns recorded speech into text. Unlike older speech-recognition tools, you can tell it in plain language what style of transcript you want: word for word or cleaned up, with speaker names, or translated. It works on finished recordings up to an hour long, not on live streams.
- Can I use Nemotron 3.5 Transcribe commercially?
- Yes. Nemotron 3.5 Transcribe 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 Transcribe open source?
- Nemotron 3.5 Transcribe 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 Transcribe on my own computer?
- Yes, on a good graphics card. The smallest variant, NVFP4, needs about 25 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Nemotron 3.5 Transcribe released?
- Nemotron 3.5 Transcribe was released by NVIDIA on October 28, 2026.
- How long a context window does Nemotron 3.5 Transcribe have?
- Nemotron 3.5 Transcribe supports a context window of up to 64K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- NVIDIA Research: Nemotron 3.5 Transcribe (28 Oct 2026) research.nvidia.com
- Nemotron 3.5 Transcribe model card (Hugging Face) huggingface.co
- Hugging Face blog: Nemotron 3.5 Transcribe and Granary v2 huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Nemotron 3.5 ASR NVIDIA | 638M | – | OpenMDW-1.1 (Permissive) | 2.93 GB Laptop or phone | |
| Parakeet TDT 0.6B v3 NVIDIA | 627M | – | CC BY 4.0 (Permissive) | 2.89 GB Laptop or phone | |
| Whisper large-v3-turbo OpenAI | 810M – 1.54B | – | MIT (Permissive) | 6 GB Laptop or phone | |
| Granite Speech 5.0 IBM | 470M – 2.31B | – | Apache 2.0 (Permissive) | 1.09 GB Laptop or phone |