NVIDIA · voice model

Nemotron 3.5 Transcribe

Nemotron 3.5 Transcribe is NVIDIA's latest voice model in our dataset, released Oct 28, 2026. Open weights (OpenMDW-1.1). NVIDIA's open-weights speech model for transcription in 40 languages, translation and speaker attribution on recordings up to an hour, released Oct 28, 2026 under the OpenMDW-1.1 license. Verified against vendor sources on Nov 1, 2026.

Reported Last verified Nov 1, 2026 · All models · JSON · Markdown

Primary source: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B. We are the fast, independent index; the vendor’s own page is the source of record for these model facts.

Nemotron 3.5 Transcribe at a glance: 65,536-token context window; takes audio, text in; returns text.

ReleasedOct 28, 2026
Context window66K65,536 tokens
LicenceOpen weightsOpenMDW-1.1

InAudioTextOutText

Key facts

Vendor-reported facts are read from the linked vendor page on the date shown above; we did not run the model. Independently tested means we ran it; no fact on this page is. Editorial is our own classification, not a vendor claim. Where a linked page does not itself state a value, the row says so. How to read provenance.

Nemotron 3.5 Transcribe: key facts with sources
Fact	Value	Label	Source
Vendor	NVIDIA	Editorial	n/a
Released	Oct 28, 2026 The model card lists Oct 28, 2026 for Hugging Face, GitHub, NGC and build.nvidia.com.	Vendor-reported	source
Status	generally available	Editorial	n/a
Context window	65,536 tokens (66K)	Vendor-reported	source
Input	audio, text	Vendor-reported	source
Output	text	Vendor-reported	source
Parameters	34B total, 4.3B active, Mamba-2–Transformer hybrid mixture-of-experts with audio and speaker encoders	Vendor-reported	source
Licence	OpenMDW-1.1 (open weights)	Vendor-reported	source

Where Nemotron 3.5 Transcribe sits in our data

Released Oct 28, 2026 (week 44 of 2026), in October 2026. Other models in our data released that month: Clover-35B-A3B-Preview, SenseNova-Loop-9B-Preview, Pika H3 Flex, Gemini 4 Carbon.
The previous NVIDIA model we track is Nemotron 3.5 Bolt E4B (4GB) (Oct 23, 2026), 5 days earlier.
Of the 65 models in our data with a published context window, Nemotron 3.5 Transcribe has the number 60 largest (66K tokens). Our data is not every model on the market.
All releases by date: AI model releases by month · new AI models this week.

Alternatives to Nemotron 3.5 Transcribe

Voice models from other vendors that share at least one of our groups with Nemotron 3.5 Transcribe. Comparable on paper, from each model’s sourced facts; we have not run them.

Voxtral 3Mistral AI · Both in: Voice · Mini Transcribe 3, per minute of audio $0.0035; Small Transcribe 3, per minute of audio $0.004; Mini Transcribe Realtime 3, per minute of audio $0.008; Mini TTS 3, per 1M characters $13; Voice agent API with Mistral Small 4 for reasoning, per minute $0.04; Voice agent API with a larger model, per minute $0.08 (see each line) · released Oct 20, 2026
Grok Voice 3.0xAI · Both in: Voice · released Oct 14, 2026
Gemini 3.8 Flash-Lite TTSGoogle · Both in: Voice · 8K context · released Sep 22, 2026
Gemini 3.8 Flash TTSGoogle · Both in: Voice · Text input $0.5; Audio output $9 (per 1M tokens) (temporary price until Dec 31, 2026; list Text input $1; Audio output $18) · released Sep 22, 2026
Grok Voice Transcribe 2.0xAI · Both in: Voice · released Sep 17, 2026
GPT-TranscribeOpenAI · Both in: Voice · Transcription $0.0045 (per minute) · released Jul 28, 2026

How Nemotron 3.5 Transcribe differs from Nemotron 3.5 Bolt E4B (4GB)

Read from the recorded facts of the two models, not from a head-to-head test. We have not run either.

Context window is 66K tokens against 64K for Nemotron 3.5 Bolt E4B (4GB).
Input is audio, text, against text for Nemotron 3.5 Bolt E4B (4GB).
Nemotron 3.5 Transcribe is a voice model; Nemotron 3.5 Bolt E4B (4GB) is a llm model.
Nemotron 3.5 Transcribe came out 5 days after Nemotron 3.5 Bolt E4B (4GB).
Facts for the other model: Nemotron 3.5 Bolt E4B (4GB).

What the sources say

These are claims from the linked pages, not our measurements.

Vendor claim (NVIDIA): best open-weights model on the English Open ASR Leaderboard, with a mean WER of 4.71% on its 9 public test sets (results captured Oct 12, 2026). We did not verify this.
One checkpoint does transcription in 40 languages, translation, speaker attribution (up to eight speakers) and audio Q&A on recordings up to one hour. It is for recorded audio, not live streaming.

What we tested

Nothing yet. We have not run Nemotron 3.5 Transcribe ourselves, so every fact on this page is Reported from the sources below. When we do test it, the scope will be stated here and in our test logs. See methodology.

Not verified, left out

We could not confirm these on a vendor page, so we do not state them.

Hosted per-minute prices: NVIDIA lists Baseten, DeepInfra and Simplismart endpoints but we did not capture their prices.
We have not run the model; all scores are NVIDIA's own.