NVIDIA Nemotron 3.5 Transcribe Review (2026)

NVIDIA released Nemotron 3.5 Transcribe on 28 October 2026, and it is the most capable free speech model the company has ever opened up. One checkpoint turns recorded audio into text in 40 languages, translates 31 of them to and from English, labels up to eight speakers, and answers questions about what was said — all steered by plain-language prompts, all running on a single GPU. The weights are free to download under an open licence, and on the public leaderboards it sits at or near the top of the open-weight field. The catch is scope: it works on recordings, not live calls, and its audio question-answering was trained in English only. This review checks NVIDIA’s claims against the company’s own launch post, the model card, and two independent leaderboards — and works out who should actually use it.

Quick answer: NVIDIA's free Nemotron 3.5 Transcribe transcribes 40 languages, translates 31, and labels eight speakers in one open-weight model — for recorded audio only.

We have not hands-on tested this model. Everything below is desk research, verified on 10 November 2026 against the sources listed at the end, and we say so plainly where a figure is the vendor’s own.

What Nemotron 3.5 Transcribe actually does

Most transcription tools do one job. Nemotron 3.5 Transcribe does four, from a single checkpoint, and the instruction you type alongside the audio picks which one you get. NVIDIA calls it a promptable speech model: the same meeting recording can come back as a plain transcript, a translation, a speaker-labelled record, or an answer to a question about the meeting.

Job	Coverage	What you get
Transcription	40 languages; recordings up to 1 hour; 65,536-token context	Verbatim or cleaned text, word-level timestamps, your choice of written form
Translation	31 languages, to and from English	The same audio rendered in English, or English audio in another language
Speaker attribution	13 languages; up to 8 speakers	Each turn tagged with who said it, labelled during generation
Audio question answering	Trained in English	Answers such as “What decisions were made in this meeting?”

The prompting is the part that makes it different from a standard speech-to-text API. NVIDIA’s launch examples show the same checkpoint answering three different instructions: “Transcribe this recording and label each speaker”, “Translate this audio into English”, and “What decisions were made in this meeting?” You can also hand it context before it listens. In a healthcare example, the prompt supplies possible terms — HbA1c, metformin — with the instruction to use them only when the audio supports them, and to preserve standard medical spelling. The model card carries recommended presets, a full prompt guide, and a task-by-language matrix, because language support differs by task: speaker labels cover 13 of the 40 transcription languages, so check the matrix before you promise a client a feature the model does not have in their language.

Under the hood it is a speech encoder feeding a hybrid Mamba–Transformer mixture-of-experts language model, with about 30 billion parameters in total and roughly 4 billion active per token. The model card counts 34 billion parameters and 4.3 billion active once the audio encoder is included, so the two figures you will see in coverage are the same model counted two ways. It ships in BF16, FP8 and NVFP4 precision, and the licence is OpenMDW-1.1, which allows commercial use.

How accurate is it? The numbers, with their caveats

Transcription accuracy is measured as word error rate (WER): the share of words a model gets wrong, so lower is better. NVIDIA reports results captured on 12 October 2026 across three public leaderboards, and the numbers are strong — with one honest wrinkle worth knowing before you quote them.

Test	Result	What it means
Hugging Face Open ASR Leaderboard, English	5.33% average WER — first among open-weight models	Just ahead of Qwen3-ASR-1.7B at 5.4%
Hugging Face Open ASR, multilingual (7 languages)	4.62% mean WER — lowest among evaluated open-weight models	The broadest language test in the set
Hugging Face Open ASR, long-form	10.98% mean WER — lowest in that comparison	Long recordings are harder; the gap widens here
Artificial Analysis AA-WER	2.81% — effectively tied with Mistral’s Voxtral Small (2.77%)	Unrounded gap is 0.04 points; NVIDIA calls it a tie for first
Artificial Analysis AA-AgentTalk	1.8% WER, versus 2.3% for the next-best open-weight model	Held-out voice-agent speech; half of the AA-WER index
David AI diarization, 21 languages	Lowest diarization error rate, 18.26%	Ahead of NVIDIA’s own dedicated Nemotron 3 Diarization at 18.38%
VoiceArena Diarization-Bench, 14 systems	Lowest cpWER, 32.4%; first in-person and online	Drops the fewest words: deletions 8.32% of reference words, versus a median of about 31%

“Transcribe is first among open-weight models on the Hugging Face Open ASR Leaderboard for English and sits on Artificial Analysis’s accuracy–active-parameter frontier.” — NVIDIA, launch post on Hugging Face, 28 October 2026

Three caveats keep these numbers honest. First, the Artificial Analysis “tie” is a rounding artefact: unrounded, Voxtral Small scores 2.77% and Nemotron 3.5 Transcribe 2.81%, and the leaderboard lists Voxtral Small first. NVIDIA’s description of a tie for first is fair as a description of practical accuracy and wrong as a description of the leaderboard order — and the real story is the efficiency: Nemotron reaches the same accuracy band with about 4 billion active parameters against Voxtral Small’s roughly 24 billion. Second, the two leaderboards measure different audio, so never merge them into one ranking; a model can lead one and trail the other. Third, every figure above is a leaderboard snapshot from October 2026 as reported by NVIDIA, not a guarantee about your audio. Noisy rooms, heavy accents and overlapping speech still degrade results — NVIDIA’s own demo uses a rapid-fire overlapping scene from the 1940 film His Girl Friday, which shows the model handling overlap, but a demo is not your meeting.

One more number matters, because it is the difference between a model that works out of the box and one that works after you tune it. In NVIDIA’s own testing, simply naming the language and alphabet in the prompt — English, Latin alphabet — cut full-benchmark WER from 30.8% to 23.8%. And in a medical case study on Eka Care’s open benchmark of recordings from Indian doctors and medical students, fine-tuning cut WER on held-out speakers by 45.8% relative to the best prompt, from 22.44% to 12.17%. The lesson is that the prompt is part of the product: budget an afternoon for prompt testing before you judge the model.

What it costs to run

The weights are free. The compute is not — but the bar is a single GPU, not a cluster, and that is the point of the design.

Precision	Hardware NVIDIA names	Best for
BF16 (full precision)	1 × H100 80 GB	Maximum accuracy on a data-centre card
FP8	L40S	Cheaper data-centre serving
NVFP4 (quantised)	RTX 5090, RTX Spark, DGX Spark, Jetson AGX Thor	A desktop, a laptop-class device, or an edge box

Throughput, on NVIDIA’s own benchmark, is high: 100 hours of audio processed in under three minutes on one RTX PRO 6000 Blackwell Server Edition GPU, using the NVFP4 quantised version on 30-second clips at high concurrency. Treat that as a vendor benchmark, not an independent measurement — your throughput will depend on clip length, concurrency and precision. If you would rather not run hardware at all, hosted endpoints are available through Baseten, DeepInfra and Simplismart; the launch material publishes no per-minute prices for them, so check the endpoint’s pricing page before you budget a project around it.

For serving, NVIDIA’s reference path is vLLM 0.32.0 with its NeMo Speech plugin, exposing an OpenAI-compatible API; local deployment with NeMo-Speech.cpp is available alongside the quantised versions, and SGLang-Omni is supported. One honest footnote from launch coverage: native vLLM integration was still under review when the model shipped, so the plugin path is the day-one route, not a native one.

How to try it in an afternoon

This is a model you deploy, not an app you sign up for, so the first test is a small one. A workable afternoon plan:

Pick your route. A hosted endpoint for the first test; self-hosting once you know the output quality suits you.
Get the checkpoint. Download nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B from Hugging Face or ModelScope, in the precision your hardware supports.
Start with ten minutes, not ten hours. One short recording, prompted with the language and the written form you want — verbatim with disfluencies, or cleaned.
Test the speaker labels on a three-person call before you trust them on an eight-person one.
Try the context trick. Supply the domain terms your audio actually contains, with the instruction to use them only when the audio supports them.
Spot-check against the recording. Names, speaker changes and answers are the three things NVIDIA itself tells users to verify — no speech model gets names right every time.

A worked example shows what the one-checkpoint design buys you. Take a one-hour board meeting with eight speakers. The same file, run three times with three prompts, returns a cleaned transcript, a speaker-labelled record with word-level timestamps, and an English translation if the meeting was not in English — plus an answer to “what was decided?” That is three tools’ worth of output from one download. The one thing the same file cannot give you is live captioning: the model is built for recorded audio, not streaming, and that is a hard limit, not a setting.

The limits that matter

Recorded audio only. This is the biggest limit and the easiest to miss. Internal audio chunking does not make it a real-time streaming service, so it cannot caption a live call or listen inside a voice agent. If your use case is live, look at a streaming model instead — TechCrunch’s launch coverage notes ElevenLabs launched a realtime model in the same month for exactly that job.

Audio question answering was trained in English. Transcription covers 40 languages, but asking questions about the audio is an English-trained capability. Do not assume it works as well in other languages.

Language support varies by task. Speaker attribution covers 13 languages, translation 31, transcription 40. The model card’s task-by-language matrix is the document of record — check it before you promise a feature in a specific language.

It is a model, not a finished product. There is no transcription app here: no editor, no billing, no support contract. You deploy it, or you pay a hosted endpoint, or you keep using a closed API that does the operations for you.

Fine-tuning needs real work. Granary v2, the open data pipeline behind the model, is genuinely open — a public manifest of 463,062 hours across 48 languages and 16 source corpora, up from 25 languages in Granary v1, with a reusable NeMo Curator pipeline and NeMo Speech fine-tuning tools. But the manifest is not the audio: source audio downloads separately under each corpus’s original licence, and the fine-tuning workflow pairs representative speech with explicit written targets. Budget data work, not just GPU time.

Nemotron 3.5 Transcribe vs Mistral Voxtral Small

The closest rival is Mistral’s Voxtral Small, and the comparison is instructive precisely because the two models land in the same accuracy band by very different routes.

	Nemotron 3.5 Transcribe (NVIDIA)	Voxtral Small (Mistral)
Artificial Analysis AA-WER	2.81%	2.77% — first on the board, unrounded
Active parameters	About 4.3 billion	About 24 billion
Hugging Face Open ASR, English	5.33% — first among open-weight models	Not published on that board
Jobs in one checkpoint	Transcription, translation, speaker labels, audio Q&A	Transcription-focused speech model
Hardware	Single GPU, down to an RTX 5090 in NVFP4	Heavier — roughly six times the active parameters

Read the table as an efficiency story, not a photo finish. Voxtral Small edges Nemotron on the Artificial Analysis index by four hundredths of a point; Nemotron leads the Hugging Face English board, where Voxtral Small has no published score, and does four jobs in one checkpoint on roughly a sixth of the active parameters. If your constraint is accuracy alone, the two are interchangeable in practice. If your constraint is hardware, cost, or doing translation and speaker labels without a second model, Nemotron is the pick. A lighter alternative on the same board is Qwen3-ASR-1.7B at 5.4% English WER — worth a look if a 1.7-billion-parameter model fits your machine and you only need transcription. Mistral, for its part, shipped Voxtral 3 in the same month, so this corner of the market is moving fast; treat every figure here as an October 2026 snapshot, not a permanent order.

Our verdict: the five-question transcription test

The original question this review set out to answer is a practical one: should this be your transcription stack? Five questions decide it.

Is your audio recorded, not live? If it is live, stop here — this is not your model.
Do you need more than a transcript? Speakers, translation or questions about the audio are where one checkpoint beats a pipeline of three tools.
Can you run one GPU, or pay a hosted endpoint? An H100 in BF16, an L40S in FP8, or an RTX 5090 in NVFP4 — or Baseten, DeepInfra or Simplismart if you would rather not run hardware.
Is your audio in the right language sets? Forty languages for transcription, thirteen for speaker labels, thirty-one for translation — and did you check the matrix?
Will you spot-check names and speaker changes? NVIDIA’s own advice, and the cheapest quality control in the whole workflow.

If you answer yes to all five, the verdict is straightforward: download it. The weights are free, the licence allows commercial use, and the accuracy is at the top of the open-weight field on two independent leaderboards. If you answer no to the first question, keep your closed API. If you answer no to the second, a smaller single-purpose model — Qwen3-ASR-1.7B at 5.4% English WER, or NVIDIA’s own 1.1-billion-parameter Parakeet — will do the same transcription job on far less hardware. The model that wins is rarely the most accurate one; it is the one whose limits match your audio.

For readers building with open-weight models more broadly, our Kimi K3.1 Pro open-weights guide covers a 2.8-trillion-parameter model with the same download-and-run shape and a very different licence catch, and our Mistral Large 4 local guide works through the hardware maths for running a trillion-parameter model at home. And if you want the sceptical frame for reading any vendor benchmark, our SemiAnalysis GDPx explainer shows what happens when the same models are scored on real work instead of leaderboards.

Frequently Asked Questions

Is NVIDIA Nemotron 3.5 Transcribe really free?

The weights are free to download from Hugging Face and ModelScope under the OpenMDW-1.1 licence, which allows commercial use. You still pay for the compute to run it, or for a hosted endpoint such as Baseten, DeepInfra or Simplismart.

Can it transcribe live calls?

No. Nemotron 3.5 Transcribe is built for recorded audio, not live streaming. It handles recordings up to one hour long, but it cannot caption a live call or listen inside a voice agent.

How many languages does it support?

Forty languages for transcription, thirty-one translated to and from English, and speaker labels in thirteen languages. Support varies by task, so check the model card’s task-by-language matrix for the language you need.

How does it compare with Mistral Voxtral Small?

On Artificial Analysis’s AA-WER index, Nemotron scores 2.81% against Voxtral Small’s 2.77% — a rounding tie, with Nemotron using about 4 billion active parameters against Voxtral Small’s roughly 24 billion. On the Hugging Face Open ASR English board, Nemotron leads the open-weight field at 5.33%.

What hardware do I need to run it?

One GPU: an H100 with 80 GB in BF16, an L40S in FP8, or an RTX 5090, RTX Spark, DGX Spark or Jetson AGX Thor in NVFP4. If you would rather not run hardware, hosted endpoints are available through Baseten, DeepInfra and Simplismart.

Sources and methodology

Claims in this article were verified on 10 November 2026 against NVIDIA’s launch post on Hugging Face (28 October 2026), the model card for NVIDIA-Nemotron-3.5-Transcribe-30B-A4B, TechCrunch’s launch coverage (28 October 2026), and Artificial Analysis’s speech-to-text leaderboard as read in independent coverage. Every benchmark figure is a leaderboard result as reported by NVIDIA, captured 12 October 2026 unless stated otherwise; the Artificial Analysis “tie” is a rounding artefact (2.77% versus 2.81% unrounded). This article is desk research — the model was not hands-on tested — and prices and endpoint availability can change after launch week.

Arva Rangwala

Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.

NVIDIA Nemotron 3.5 Transcribe Review: Free Speech AI in 40 Languages