What is NVIDIA Nemotron 3.5 Transcribe?
Nemotron 3.5 Transcribe is an offline speech model from NVIDIA, released on Hugging Face, NGC and build.nvidia.com on 28 October 2026. It pairs a 600M-parameter ASR encoder and a 99M-parameter speaker encoder with the Nemotron 3.5 Lightning hybrid Mamba-Transformer mixture-of-experts language model: 34 billion parameters in total, about 4.3 billion active per token including the audio encoder.
One checkpoint handles transcription in 40 languages, verbatim or cleaned-up text, speaker-attributed transcripts with word timestamps, translation to and from English, language identification, vocabulary hints in the prompt and English audio question answering. It takes recordings of up to one hour and is not a streaming model.
The weights are under the OpenMDW-1.1 licence, which permits commercial use, in BF16, FP8 and NVFP4 builds. NVIDIA also released Granary v2, 463,062 hours of speech data in 48 languages. DeepInfra hosts it at $0.002 a minute of audio.
How much does NVIDIA Nemotron 3.5 Transcribe cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | OpenMDW-1.1, commercial use allowed |
| DeepInfra API | $0.002 per audio minute | Up to 60 minutes per request |
Open weights under OpenMDW-1.1. DeepInfra: $0.002 per minute of audio ($0.12 an hour); OpenRouter routes to DeepInfra at about the same rate.
Is NVIDIA Nemotron 3.5 Transcribe worth it?
Nemotron 3.5 Transcribe is the strongest open transcription model on independent tests, and it is the first open one that does most of a meeting-notes pipeline in a single call. Artificial Analysis puts it at 2.8% on its AA-WER index, tied for first among open-weight models with Mistral's Voxtral Small (2.77% to its 2.81%) while using about a sixth of Voxtral's 24B active parameters. On Hugging Face's Open ASR Leaderboard it ranks first among open models at 5.33% average English WER, ahead of Qwen3-ASR-1.7B at 5.4.
The distinctive part is the prompt. Transcription style, vocabulary hints, speaker tags and translation are all chosen in the instruction, and NVIDIA's own multi-speaker figures (15.80 mean cpWER across AMI, CALLHOME, Mixer 6 and NOTSOFAR-1 full sessions, against 16.83 for the next best published system) are self-reported but consistent with the outside rankings. Domain adaptation is cheap: on Eka Care's medical speech set, WER fell from 30.8% to 23.8% with a better prompt and to 11.8% after about 14 GPU-hours of fine-tuning on one GB300.
The limits are clear from the card. It is offline only, so live captions and voice agents that need partial results still want a streaming model such as Parakeet. Language support is uneven: 40 languages for transcription, 31 for translation and 13 for speaker labels, with Mandarin, Arabic and Bengali rated only "supported". Audio question answering was trained in English. Closed models remain ahead overall; Gemini 3.5 Transcribe sits at 2.60% on the Open ASR board.
Self-hosting needs one 80GB H100 for the BF16 weights. NVIDIA's fastest serving numbers, 2,238 seconds of audio per second on an RTX PRO 6000, come from its NIM build; the open vLLM path shipped on release day is, by NVIDIA's own estimate, about half as fast. At DeepInfra's price, transcription costs a third of OpenAI's historical Whisper API rate, so the hosted route will be the sensible default for most teams below a few thousand hours a month.
Where it earns its place
A data engineer at a 300-seat contact centre
Transcribe and attribute 10,000 hours of recorded calls a month
On DeepInfra at $0.12 an hour that is about $1,200 a month, with <spk:N> speaker tags and word timestamps in one response. Product names go in the prompt as context terms rather than a separate biasing step.
An ML engineer at a health-tech startup
Adapt transcription to clinical vocabulary
NVIDIA's Eka Care example cut WER from 30.8% to 11.8% with about 245 hours of data and 14 GPU-hours on one GB300. General English slipped slightly (LibriSpeech clean 1.54% to 1.69%), so test both.
A localisation producer at a European media company
Produce English transcripts and translations of interviews in Czech, Polish or German
Those languages are rated "leading" for transcription and translation in both directions. One-hour recordings fit in a single request; names and places still need spot-checking against the audio, as NVIDIA advises.
Before you commit
- Setup
- Hosted: one OpenAI-style call to DeepInfra or OpenRouter (nvidia/nemotron-3.5-transcribe). Self-hosted: vLLM 0.32.0 with the NeMo Speech plugin on one H100 80GB, or SGLang-Omni 0.2.0; Transformers works from source.
- Integrations
- vLLM, SGLang-Omni, Hugging Face Transformers, GGML on RTX 5090 and RTX PRO 6000, NeMo Speech for fine-tuning; hosted on DeepInfra, Baseten and Simplismart.
- Data and privacy
- OpenMDW-1.1 permits commercial use. NVIDIA says no user interactions with deployed products were used in training. DeepInfra says it does not store audio or transcripts after a request.
- Cost at scale
- 1,000 hours of audio costs $120 on DeepInfra, which bills every minute including silence. Self-hosting needs one 80GB H100 (BF16) or an L40S 48GB (FP8).
- Limits
- Up to one hour and 65,536 tokens per request, up to eight speakers; offline only; audio QA tested in English only; the card notes it is not guaranteed to generalise beyond the one-hour training limit.
What’s good
- Tied first among open models on Artificial Analysis at 2.8% AA-WER, with about 4B active parameters
- Speaker labels, word timestamps, translation and vocabulary hints in one checkpoint
- $0.002 per audio minute on DeepInfra
- OpenMDW-1.1 licence and published fine-tuning recipes in NeMo Speech
What’s not
- Offline only; no streaming partial results
- Speaker labels in 13 languages, translation in 31
- Fastest serving path is NIM-only for now; open vLLM is about half the speed
- BF16 weights need an 80GB GPU
Best for
Teams transcribing recorded calls, meetings or media at volume who want speaker labels and domain vocabulary without stitching several models together.
NVIDIA Nemotron 3.5 Transcribe alternatives
NVIDIA Nemotron 3.5 Transcribe: common questions
- Is NVIDIA Nemotron 3.5 Transcribe worth it?
- An open speech model that matches the best open rivals on accuracy and folds diarization, translation and custom vocabulary into one prompt, for $0.12 an hour hosted. We rate it 4.5 out of 5.
- How much does NVIDIA Nemotron 3.5 Transcribe cost?
- Open weights under OpenMDW-1.1. DeepInfra: $0.002 per minute of audio ($0.12 an hour); OpenRouter routes to DeepInfra at about the same rate.
- What are the downsides of NVIDIA Nemotron 3.5 Transcribe?
- Offline only; no streaming partial results; speaker labels in 13 languages, translation in 31; fastest serving path is NIM-only for now; open vLLM is about half the speed; bF16 weights need an 80GB GPU.
- Who is NVIDIA Nemotron 3.5 Transcribe best for?
- Teams transcribing recorded calls, meetings or media at volume who want speaker labels and domain vocabulary without stitching several models together.
- How much does NVIDIA Nemotron 3.5 Transcribe cost for a team or in production?
- 1,000 hours of audio costs $120 on DeepInfra, which bills every minute including silence. Self-hosting needs one 80GB H100 (BF16) or an L40S 48GB (FP8).
- Is NVIDIA Nemotron 3.5 Transcribe safe for company data?
- OpenMDW-1.1 permits commercial use. NVIDIA says no user interactions with deployed products were used in training. DeepInfra says it does not store audio or transcripts after a request.
- What are the best NVIDIA Nemotron 3.5 Transcribe alternatives?
- The closest alternatives we have reviewed are Voxtral 3 (Transcribe and TTS), Parakeet TDT 0.6B v3, Whisper large-v3-turbo and MOSS-Transcribe-Diarize 0.9B.
Answers checked Oct 31, 2026.
Make NVIDIA Nemotron 3.5 Transcribe? Show this rating on your site
