Nemotron 3.5 Transcribe: 40 languages, open weights

What’s new, in 20 seconds

Nvidia released Nemotron 3.5 Transcribe on Hugging Face on October 28, 2026: one open-weight model for transcription in 40 languages, plus translation, speaker-attributed transcripts, timestamps and audio Q&A. It handles recorded audio up to one hour, not live streams.
The weights are under OpenMDW-1.1, the license Nvidia also used for Nemotron 3.5 Super VL, which allows commercial use with no use or output restrictions. The model card gives about 34B total and 4.3B active parameters, including the audio encoder; the name and Nvidia's blog say 30B.
Nvidia reports a 4.71% mean English word error rate on the card's nine-set comparison. The Open ASR Leaderboard lists 5.33%, which Nvidia says ranks first among open-weight models. Artificial Analysis lists 3.4% AA-WER in its table and 2.8% in its FAQ. The card's deployment target is one H100 80GB GPU.

Contents

Nemotron 3.5 Transcribe is Nvidia's new open-weight speech recognition model, released on Hugging Face on October 28, 2026. One checkpoint transcribes 40 languages, translates a subset to and from English, labels who is speaking, adds word timestamps and answers questions about a recording. It comes under OpenMDW-1.1, the same permissive license Nvidia used for Nemotron 3.5 Super VL and Bolt on October 15, 2026, and it is a voice AI model anyone can download and use commercially. It is built for recorded audio, not live transcription, and most accuracy figures so far come from Nvidia.

This story is based on the Nemotron 3.5 Transcribe model card on Hugging Face, Nvidia's October 28 Hugging Face blog post on Transcribe and its Granary v2 dataset, the OpenMDW-1.1 license text, the Hugging Face card for Nvidia's earlier Nemotron 3.5 ASR streaming model, the Hugging Face Open ASR Leaderboard and Artificial Analysis' speech-to-text leaderboard, checked October 29, 2026.

The short version

What it is: an offline, multitask speech model. Nvidia's card lists transcription, translation, speaker-attributed transcription, timestamps, language identification and audio question answering.
Languages: 40 for transcription. Nvidia's blog says translation covers 31 of them to or from English, and speaker attribution 13.
Size: about 34B total and 4.3B active parameters, including the audio encoder, per the model card. The repo name and Nvidia's blog say 30B with about 4B active.
License: OpenMDW-1.1, which allows commercial use with no restrictions on use or outputs.
Accuracy: Nvidia's figures are vendor-reported. The Hugging Face Open ASR Leaderboard lists it at 5.33% average English WER, which Nvidia says ranks first among open-weight models.
Hardware: one Nvidia GPU. The card's deployment target is a single H100 80GB.

What is Nemotron 3.5 Transcribe?

It pairs an audio encoder with a hybrid Mamba-Transformer mixture-of-experts language model, so only a fraction of its weights work on each token. Nvidia's blog says speaker cues from its Sortformer diarization model are fed to the language model alongside the audio, so speaker labels are written as part of the transcript rather than added afterward.

According to the model card, it accepts 16 kHz single-channel audio up to one hour long, with a 65,536-token context. Nvidia's blog says it handles up to eight speakers. You steer it with prompts: the card gives recommended prompts for verbatim or cleaned-up transcripts, speaker attribution, code-switching and language ID.

It is not Nvidia's first Nemotron 3.5 speech model. Nemotron 3.5 ASR, released on June 4, 2026, is a 600M-parameter streaming model that also covers 40 language-locales under OpenMDW-1.1, built for live captions and voice agents with chunk sizes from 80 ms. Transcribe is the much larger offline sibling. The card is explicit that "internal audio chunking does not make it a real-time streaming ASR service."

Which 40 languages does it support?

The model card's language table sorts them into three tiers:

Leading: Czech, Dutch, English (US, India and UK), French (France and Canada), German, Hindi, Italian, Korean, Polish, Portuguese (Portugal and Brazil), Russian, Slovak and Spanish (US and Spain).
Production-ready: Bulgarian, Croatian, Danish, Estonian, Finnish, Hungarian, Japanese, Latvian, Lithuanian, Romanian, Slovenian, Swedish, Thai, Turkish and Ukrainian.
Supported: Arabic, Bengali, Mandarin Chinese, Greek, Hebrew, Indonesian, Maltese, Marathi, Tagalog, Tamil, Telugu and Vietnamese.

That's 40. The table also has a 41st row, Norwegian, marked "Supported" but with no transcription checkmark, so check the card before relying on it for Norwegian. The card says the audio-understanding, reasoning and instruction-following features were evaluated only in English "and may not work reliably in other languages."

What license is Nemotron 3.5 Transcribe under?

OpenMDW-1.1. The model card says use "is governed by the OpenMDW License Agreement, version 1.1" and that the model "is ready for commercial or non-commercial use." The license itself grants permission "to deal in the Model Materials without restriction." Its conditions are short:

Keep a copy of the license and the copyright notices when you redistribute.
No acceptable-use policy, and no restrictions on outputs.
Your rights end if you sue claiming the materials infringe a patent or copyright, unless you're responding to a suit brought against you.

Third-party components bundled with the release, such as vLLM patches, carry their own licenses, per the card. Nvidia's blog says the Granary v2 training dataset follows the licenses of its source corpora. For how OpenMDW compares with Apache 2.0 and the revenue-gated licenses, see our open-weight licenses comparison.

How big is it, and what hardware does it need?

Item	Nemotron 3.5 Transcribe
Parameters	About 34B total, 4.3B active, including the audio encoder (model card); "30B" with about 4B active (blog and repo name)
Weights	BF16, plus FP8 and NVFP4 versions
Input	16 kHz mono, up to one hour, up to eight speakers
Deployment target	1× H100 80GB
Other listed GPUs	A100, H200, B200, GB200, DGX Spark, RTX 5090, RTX PRO 6000; L40S (FP8 only)
Runtimes	vLLM 0.32.0, Hugging Face Transformers, SGLang-Omni 0.2.0, GGML on some GPUs
Streaming	No
License	OpenMDW-1.1

Table 1The card and Nvidia's own blog disagree on total size, and the card gives no memory figure for any precision. Source: Nemotron 3.5 Transcribe model card and Nvidia's Hugging Face blog, October 28, 2026.

Last verified: October 29, 2026

Nvidia's blog says Transcribe "runs on a single NVIDIA GPU, from Jetson AGX Thor and DGX Spark to Blackwell data-center GPUs." The card lists Linux and Windows (through WSL2, or natively with GGML). Neither page states how much GPU memory each precision needs.

On speed, Nvidia's blog says one RTX PRO 6000 Blackwell Server Edition running NVFP4 processes 100 hours of audio in under three minutes, measured on 30-second clips at high concurrency. The card gives rough batched figures of 260 to 410 times real time on one H100 in BF16, depending on the decoding setup, and calls them estimates. These are Nvidia's numbers.

What word error rates does Nvidia report?

Nvidia's own figures are vendor-reported: Nvidia ran them and published them on its card and blog. The leaderboard listings are third-party. Word error rate (WER) is the share of words transcribed wrongly; lower is better.

English (Open ASR Leaderboard sets): the model card gives a 4.71% mean across nine datasets, against 4.94% for Alibaba's Qwen3-ASR-1.7B, the next-best open-weight model in its comparison. Its per-dataset results run from 1.09% on LibriSpeech test-clean to 11.00% on URGENT24. The card says its results were captured on October 12, 2026.
The leaderboard itself: Nvidia's blog says Transcribe "is first among open-weight models on the Hugging Face Open ASR Leaderboard for English," at 5.33% average WER and an RTFx (speed relative to real time) of 462.09. The leaderboard, which added the model on October 28 according to its changelog, shows the same 5.33% and 462.09 as of October 29, behind at least eight other listed models, including ElevenLabs' Scribe v2 at 5.24%. Its average also counts two private test sets, scripted (3.58%) and conversational (12.73%), that the card's nine-set mean leaves out; averaging all eleven columns gives 5.33% (9to5AI's arithmetic). The leaderboard doesn't flag which models are open-weight, so the "first among open-weight models" ranking is Nvidia's.
Multilingual (seven languages): 4.62% mean WER, against 6.49% for Qwen3-ASR-1.7B on the card.
Long recordings: 10.98% mean WER, against 13.37% for Nvidia's Parakeet-TDT-0.6B-v3 on the card.
Artificial Analysis: Nvidia's blog says Transcribe ties for first among open-weight models at 2.8% on Artificial Analysis' AA-WER index, in a chart dated October 22. Artificial Analysis' own speech-to-text page, checked October 29, gives two figures: its table lists Transcribe, served through DeepInfra, at 3.4% AA-WER, while its FAQ lists it at 2.8%, level with Mistral's Voxtral Small, among the most accurate open-weights models. The page doesn't explain the difference.

Two caveats. Nvidia chose the comparison models, and its tables compare against small open models such as Qwen3-ASR-1.7B and Parakeet, not against the API models that top Artificial Analysis' index, such as Alibaba's Fun-Realtime-ASR-preview (1.7%) and Microsoft's MAI-Transcribe-2 (2.0%), as listed on October 29, 2026. For context, our Voxtral 3 story covers Mistral's API-only transcription models, which sit near the top of that index.

What's confirmed, reported and uncertain

Confirmed (model card, license text): the October 28 release; 40 transcription languages and the tier list; offline, not streaming; one-hour, 16 kHz input; BF16, FP8 and NVFP4 weights; OpenMDW-1.1; the hardware and runtime tables; 34B total and 4.3B active on the card.
Vendor-reported: every WER and speed figure on the card and blog, including the "first among open-weight models" Open ASR ranking; the 31-language translation and 13-language speaker-attribution counts (blog).
Third-party listings: the Open ASR Leaderboard's 5.33% and 462.09 RTFx; Artificial Analysis' 3.4% (table) and 2.8% (FAQ) AA-WER, checked October 29.
Uncertain: the total parameter count (34B on the card, 30B in the blog); which of Artificial Analysis' two figures is current; GPU memory needs; Norwegian support.

For another recent open voice release, see Hurston-1, an Apache 2.0 text-to-speech model, and for a smaller speech-to-text experiment, our EmbeddingGemma 2 transcription story.