Skip to content
TrustList
All launches
AI model launch

Nemotron 3.5 Transcribe 30B-A4B

NVIDIA's open-weight speech model, released 28 October 2026, transcribes and translates recordings of up to an hour in 40 languages, with speaker labels and word timestamps, on one H100 GPU.

Launched 2026-10-28AI model

About this launch

Nemotron 3.5 Transcribe turns recorded speech into text. NVIDIA released it on 28 October 2026 on its build.nvidia.com catalogue, GitHub, Hugging Face and NGC, under the OpenMDW-1.1 licence, which its model card says allows commercial use. It works offline on files rather than on live streams, and handles recordings up to one hour long.

## One model for several transcription jobs

The model is a speech-augmented language model: an audio encoder feeding a mixture-of-experts language model of about 30 billion parameters, of which about 4.3 billion are active for each token. An instruction in the prompt selects the job. The model card lists verbatim or written-style transcription, transcription that uses supplied context such as names and terms, transcripts labelled by speaker, rich transcription, word timestamps, translation of speech, language identification and answering questions about a recording. It covers 40 languages and has a 65,536-token context. A configurable reasoning mode is meant for the question-answering tasks; NVIDIA recommends switching it off for plain transcription.

## Running it on one GPU

Weights come in BF16, FP8 and NVFP4 precision, and NVIDIA says the model runs on one H100 80 GB GPU. The release ships with a vLLM integration in its NeMo Speech toolkit. NVIDIA's faster serving path is a NIM container: on one RTX PRO 6000 Blackwell Server Edition GPU it reports transcribing up to 2,238 seconds of audio per second and returning a 30-second transcript in 0.13 seconds. The card says those serving optimisations are not yet in the open-source release and that the vLLM integration shipped on release day is about half as fast.

## What NVIDIA reports

On its own tests, captured on 12 October, NVIDIA reports a mean word error rate of 4.62 on seven languages of the Multilingual Open ASR Leaderboard, against 6.49 for Qwen3-ASR-1.7B, and a mean speaker-attributed error rate of 15.80 across full AMI, CALLHOME, Mixer 6 and NOTSOFAR-1 sessions. These are NVIDIA's own evaluations and are not stored as scores here. The card notes that the training data included personal data and describes the filtering NVIDIA applied.

## Key facts

**Maker:** NVIDIA **Released:** 28 October 2026 **Size:** about 30B parameters, about 4.3B active **Languages:** 40 **Input and output:** audio and text in, text out, optional word timestamps **Context:** 65,536 tokens; recordings up to one hour **Licence:** OpenMDW-1.1 **Board family:** Nemotron 3.5, after the Embed 8B and Rerank 8B models

Not yet independently verified: Not yet independently verified. The release date and every figure here come from NVIDIA's own model card; we found no independent test of the model yet. We will update this when it can be confirmed, and remove this note.

Discussion

Questions & answers about Nemotron 3.5 Transcribe 30B-A4B.

No questions yet. Be the first to ask about Nemotron 3.5 Transcribe 30B-A4B.