Nvidia releases promptable Nemotron 3.5 Transcribe model

Nvidia has released Nemotron 3.5 Transcribe, a new open-weight automatic speech recognition model designed for recorded audio. Launched on October 28, 2026, the model allows developers to provide specific instructions before processing audio files. It currently leads the Hugging Face Open ASR leaderboard for English with a word error rate of 5.33%.

What happened

The core innovation in this release is the model’s ability to accept prompts. Unlike traditional transcription engines that simply convert sound to text, Nemotron 3.5 Transcribe lets users set the context beforehand. A developer can specify that an audio file contains a medical discussion in English and provide a list of specialized terminology. The model uses this context to reduce errors and improve relevance.

Users can also instruct the model on output formatting. Requests can include labeling individual speakers, producing a verbatim transcript versus a cleaned-up version, or defining how numbers should appear in the text. Nvidia bundles these capabilities, including translation features, into a single checkpoint. This consolidation aims to simplify the architecture for applications that require multiple speech processing tasks.

The model supports transcription for 40 languages and can translate 31 of those languages to or from English. Speaker attribution, which identifies who is speaking, works for 13 languages and can distinguish between up to eight different voices. However, the Audio Question and Answering feature was trained only in English. The system handles recordings up to one hour in length but does not support live streaming audio.

How it works

Nemotron 3.5 Transcribe is a 30 billion parameter mixture-of-experts model. In this architecture, only a subset of the total parameters is active for any given input. For this model, there are 4 billion active parameters. This design allows it to rival the accuracy of larger models while maintaining lower computational demands during inference. For example, it achieves a 2.81% word error rate on Artificial Analysis benchmarks, comparable to Mistral’s Voxtral Small, which uses 24 billion active parameters.

Developers can deploy the model using vLLM or run it locally with NeMo-Speech.cpp. Nvidia states that the model runs on a single GPU, supporting hardware from the Jetson AGX Thor and DGX Spark to Blackwell data-center GPUs. Because it is an open-weight model under the OpenMDW-1.1 license, users can fine-tune it for specific domains. Nvidia reported that fine-tuning for medical transcription reduced the word error rate from 22.44% to 12.17% on held-out speakers, a relative reduction of 45.8%. This adaptation required approximately 14 GPU-hours on a single GB300 unit.

To support customization, Nvidia is also releasing Granary v2, a data curation pipeline built with NeMo Curator. This tool includes public manifests covering roughly 463,062 hours of audio across 48 languages and 16 source corpora. Developers can use this pipeline to prepare their own recordings for fine-tuning, though these figures represent the dataset release rather than the full training mixture used for the base model.

Key details

• Model architecture: 30 billion parameter mixture-of-experts with 4 billion active parameters.
• Performance: Top rank on Hugging Face Open ASR leaderboard for English with 5.33% word error rate.
• Language support: Transcribes 40 languages; translates 31 languages to or from English.
• Speaker handling: Attributes speakers in 13 languages and distinguishes up to eight voices.
• Deployment: Runs on a single GPU using vLLM or NeMo-Speech.cpp; supports Jetson AGX Thor to Blackwell GPUs.
• License: Open-weight under OpenMDW-1.1, allowing for fine-tuning and commercial use.

Why it matters

For software engineers building voice-enabled applications, the ability to prompt the transcription model reduces the need for complex post-processing scripts. Traditionally, developers had to build separate systems to clean up transcripts, format numbers, or identify speakers. By handling these tasks within the model via prompts, the development workflow becomes simpler and more maintainable. The single checkpoint approach also reduces the infrastructure overhead of managing multiple specialized models.

The open-weight nature of the model is significant for industries with strict data privacy requirements, such as healthcare and finance. Organizations can fine-tune the model on their private data without sending sensitive audio to third-party APIs. The reported reduction in error rates after fine-tuning demonstrates that domain-specific adaptation is effective, making high-accuracy transcription accessible for specialized use cases without requiring massive computational resources for training from scratch.

What you can do

• Download the Nemotron 3.5 Transcribe weights from Hugging Face under the OpenMDW-1.1 license.
• Test the model locally using NeMo-Speech.cpp on compatible single-GPU hardware.
• Use the Granary v2 pipeline to curate and prepare your own audio datasets for fine-tuning.
• Experiment with prompt engineering to define context, speaker labels, and output formatting for your specific domain.
• Evaluate the model’s performance on your target languages, noting that Audio Q&A is currently English-only.
• Consider fine-tuning if you work in specialized fields like medicine or law to significantly reduce word error rates.