# NVIDIA opens Nemotron 3.5 Transcribe: one 4.3B-active model for 40-language speech, speakers and translation

> OnPrem World · https://onpremworld.com/en/news/2026-10-nemotron-3-5-transcribe · CC BY 4.0

Published: 2026-10-28 · Models

NVIDIA released Nemotron 3.5 Transcribe on October 28 as an open-weight, promptable speech model under the OpenMDW-1.1 license. One checkpoint transcribes 40 languages, translates 31 of them to or from English, attributes speakers in 13, labels up to eight speakers and answers questions about audio, on recordings up to one hour. It activates about 4.3 billion parameters and comes in BF16, FP8 and NVFP4 versions on Hugging Face. It is an offline model, not a real-time streaming service.

## Why it matters locally

Meeting transcription, speaker labels and translation can now stay on your own machine with one model instead of several, and the 21.6 GB NVFP4 weights are small enough to consider for a workstation GPU.

## For the market

NVIDIA says it has 80% fewer active parameters than the top open-weight model on Artificial Analysis, where it ranks second, and Hugging Face's Open ASR Leaderboard has it first among open-weight models for English. Bundling transcription, diarization and translation could squeeze specialised speech pipelines, and it also gives buyers another reason to choose NVIDIA hardware, since the model is built for NVIDIA GPUs.

## In plain words

Record a meeting, then tell this AI what you want: a transcript with names of who spoke, a translation, or a list of decisions. It runs on your own computer, like a secretary who never takes the recording out of the building, but it works on finished recordings rather than live speech.

## For practitioners

Weights are 67.5 GB in BF16, 34.7 GB in FP8 and 21.6 GB in NVFP4; the card lists NVFP4 for the RTX 5090, and vLLM 0.32.0 with the NeMo Speech main branch as the runtime on Linux or WSL2, with native Windows through GGML on supported GPUs. NVIDIA's 2,238x real-time figure comes from its NIM, and the open vLLM integration is expected to be about two times slower. Input is 16 kHz mono, and Mandarin is marked 'Supported', meaning it may need fine-tuning.

## Source

[NVIDIA Nemotron 3.5 Transcribe model card (Hugging Face)](https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B)

## On this site

- Hardware: [DGX Spark](https://onpremworld.com/en/hardware/nvidia-dgx-spark.md)
- Hardware: [RTX Spark (N1X) laptops & compact desktops](https://onpremworld.com/en/hardware/nvidia-rtx-spark-n1x.md)
- Hardware: [GeForce RTX 5090](https://onpremworld.com/en/hardware/rtx-5090.md)

[All news](https://onpremworld.com/en/news.md)

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.