Nemotron 3.5 Transcribe: a 34-billion-parameter open speech model

On 28 October 2026 Nvidia released Nemotron 3.5 Transcribe, an open-weight model that handles transcription, translation, speaker attribution and questions about a recording in one place. The OpenMDW-1.1 license has no revenue cap, and at 34 billion parameters the model is unusually large for speech recognition. This post keeps two questions apart: does it fit hardware like ours, and does the accuracy Nvidia reports hold up?

Nemotron 3.5 Transcribe is an open speech model from Nvidia for audio. It has 34 billion parameters, 4.3 billion of them active, and processes recordings of up to one hour offline. It transcribes 40 languages, translates 31 languages to and from English, attributes speakers in 13 languages and answers English questions about what it heard. The weights are released under the OpenMDW-1.1 license.

One model, several tasks

The model card on Hugging Face describes a model that normally takes several separate programs. The prompt selects the task: transcription in 40 languages, translation between English and 31 other languages, transcription with speaker labels (13 languages, German among them) and answers to questions about the audio. According to the card, question answering was evaluated in English only, so anyone who needs it in another language has to test that themselves.

The model works offline. The card says outright that its internal 30-second chunking is not streaming. It was trained on recordings of up to one hour with a context of 65,536 tokens. That covers meetings and lectures but not a full working day in one file. One note on the name: the repository is called 30B-A4B, while the card gives 34 billion parameters in total because the audio encoder is counted, and 4.3 billion active. The Hugging Face API reports 33,697,781,029 parameters, which matches the 34 billion.

The language part is a Mamba-2 and Transformer hybrid with mixture-of-experts, next to a Transformer encoder for speech recognition and a Sortformer encoder for speakers. The card names Nemotron 3.5 Lightning 30B-A3B as the base. From the same model family, our post on Nemotron 3.5 Embed, Rerank and Parse covers the retrieval models for RAG. This one is about audio instead: speech recognition, hardware needs and license.

The numbers and where they come from

Every value below is on the model card or in a repository file listing. We read them through the Hugging Face API on 5 November 2026; nothing was run on our own hardware.

On the training audio the card says the first figure counts each indexed record once per source. The same recording can appear in several sources, so these are not deduplicated recording hours. Granary v2 is far smaller than the training mix, at 463,062 hours in 48 languages from 16 corpora.

Which hardware it runs on

The card names one H100 with 80 GB as the target and lists precision and runtime per card. From the card's table, as of 5 November 2026:

• H100 and H200: BF16 and FP8 with vLLM and Transformers.
• A100 80 GB: BF16. L40S with 48 GB: FP8 only.
• RTX PRO 6000 Blackwell Server Edition: BF16, FP8 and NVFP4, plus GGML and SGLang-Omni.
• RTX 5090: NVFP4 only, with vLLM, Transformers and GGML.
• Also listed: B200, GB200, GB300, DGX Spark and Jetson Thor.

The 67.5 GB of BF16 weights do not fit on an RTX 5090 with 32 GB, and neither does the 34.7 GB FP8 repository. The NVFP4 files add up to 21.6 GB, below the card size. What context and cache add on top is not stated for the RTX 5090, and we have not measured it. The FP8 and NVFP4 repositories are publicly reachable (API on 5 November 2026: HTTP 200). More posts on GPUs and memory needs of local models are collected in our topic hub on local AI and GPU hardware.

For scale, our post on Parakeet Redux covers a recognition model squeezed to 177.8 MB by Moondream, derived from Nvidia's Parakeet TDT 0.6B v3. We only read about it and did not install it. Against Nemotron's 67.5 GB that is about two orders of magnitude. Parakeet Redux recognizes 25 languages, Nemotron 40 and it can also translate.

What the accuracy is worth

The model card claims the best results among open-weight models on several tests (measured 12 October 2026):

The values in text: English, mean of nine Open ASR leaderboard tests, 4.71 against 4.94 percent for Qwen3-ASR-1.7B. Long recordings 10.98 against 13.37 for Parakeet-TDT-0.6B-v3. Seven languages of the multilingual leaderboard 4.62 against 6.49. Multiple speakers 15.80 against 16.83 (cpWER, word errors including speaker attribution).

Four caveats:

1. These are Nvidia's figures and we did not re-measure them.
2. In the English comparison a model with 34 billion parameters faces one with 1.7 billion. The gap is 0.23 percentage points.
3. The Open ASR leaderboard entries that ship with the model card name the card itself as their source in open_asr_leaderboard.yaml. We did not read the leaderboard itself.
4. Artificial Analysis uses its own method. On 5 November 2026 its table row for Nemotron 3.5 Transcribe at provider DeepInfra read 3.4 percent AA-WER. The summary text of the same page names Voxtral Small from Mistral and Nemotron 3.5 Transcribe as the best open models at 2.8 percent each. So the 3.4 percent applies to the DeepInfra row only, and the page contradicts itself here. It shows no lead for the vendor, and a deficit is not established either.

For German the card gives 2.73 percent word error on FLEURS and 2.98 percent on CoVoST2. In the language matrix German sits at the top tier ("Leading"), including translation and multiple speakers. That is the vendor's own grading. We have no measurement on our own recordings.

On throughput the card cites up to 2,238 seconds of audio per second on an RTX PRO 6000 with NVFP4 and the draft model, and 0.13 seconds for a 30-second recording. Nvidia adds that these optimizations of the Nemotron Transcribe NIM are not yet part of the open release. The bundled vLLM integration is expected to be roughly half as fast.

The license is the real news

The license is in the repository's license file. According to openmdw.org, OpenMDW was introduced by the Linux Foundation and the PyTorch Foundation. The text grants free-of-charge permission to deal in the model materials without restriction. It holds two conditions: anyone who distributes the material keeps the license text and the notices of origin, and anyone who sues claiming the model materials infringe patents or copyright loses their rights, unless the suit answers one brought against them. The license puts no restrictions on the outputs of the model.

What matters is what is missing: a revenue cap. The LFM Open License from Liquid AI, for example, bars commercial use from 10 million US dollars in annual revenue without contacting the vendor, as we read in the license text for our post on LFM3-Audio in llama.cpp. Anyone building speech recognition into a service needs no such clearance under OpenMDW-1.1. This is not legal advice. The license text also counts documentation and related data as part of the model materials; what that means for the training data we have not checked. For fine-tuning, the model card points to recipes for NeMo Speech.

What is checked and what is missing

We read the model card, the license text, the repository file lists and the Artificial Analysis page. Accuracy, throughput and behavior on German with our own material are not checked, and the model does not run here. A sensible test has two steps: first find out whether the NVFP4 variant starts on a 32 GB card with a realistic context, then run real meeting recordings through it and compare against the current solution. Anyone who wants to run such models in-house will find the building blocks in our service for company-owned AI systems.

In short: the license is a gain for any business that uses speech models commercially. The size is a hurdle for small cards. The accuracy is strong by Nvidia's numbers and only partly confirmed independently: Artificial Analysis shows 3.4 percent for the DeepInfra row and 2.8 percent, level with Voxtral Small, in its summary text.

Updated on 5 November 2026

We re-read the post against its sources and corrected it. Artificial Analysis shows 3.4 percent for Nemotron in its table (DeepInfra row) but 2.8 percent, level with Voxtral Small, in the page's summary text; both are now stated and the ranking we first claimed is removed. The FP8 and NVFP4 repositories are reachable, the BF16 file size is given in gigabytes (67.5) instead of gibibytes, and the statements on the 178 MB model, on the patent clause and on other licenses are fixed.

Further sources

• Hugging Face: NVIDIA-Nemotron-3.5-Transcribe-30B-A4B: model card with hardware table, benchmarks and language matrix, read on 5 November 2026
• Hugging Face: LICENSE in the repository: text of the OpenMDW-1.1 license
• Nvidia on Hugging Face: Nemotron Transcribe and Granary: post of 28 October 2026 on the data pipeline
• Hugging Face: Granary v2: dataset card with 463,062 hours in 48 languages
• OpenMDW: project site of the license, Linux Foundation
• Artificial Analysis: Speech to Text: independent leaderboard, read on 5 November 2026
• Video: none is embedded; the post rests only on the text sources above.