NVIDIA released Audex-30B — a unified audio-text MoE model based on Nemotron-Cascade-2

NVIDIA released Audex-30B on July 7, 2026 — a unified neural network for audio and speech on a Mixture of Experts architecture, which combines five types of audio tasks in a single model without significant loss of text abilities of the base system.

What Audex combines in one model

Before Audex, each audio task required a separate specialized solution: one system for speech recognition, another for translation, a third for synthesis. NVIDIA combined them in a single MoE architecture:

• Understanding audio content — analysis of the content of audio files
• Automatic Speech Recognition (ASR)
• Speech translation to another language
• Speech synthesis — text-to-speech (TTS)
• Audio generation

The Mixture of Experts architecture allows the model to activate only part of the parameters depending on the specific task. With a total volume of 30 billion parameters, approximately 3 billion are activated during inference — hence the A3B suffix in the full name. This approach reduces computational load during inference compared to a dense model of similar total size.

Why preserving text abilities is important

Adding a new modality to an already trained language model is traditionally accompanied by "catastrophic forgetting" — degradation of the original text abilities under the influence of new training. This is one of the key engineering challenges when developing multimodal systems.

NVIDIA built Audex on the basis of Nemotron-Cascade-2 — its own text base with strong language performance. The company claims that regression on text benchmarks turned out to be minimal: audio capabilities were added without noticeable loss of quality in working with text.

The MoE architecture plays a key role here: audio and text "experts" within the model are partially isolated from each other, which reduces mutual influence of modalities during training. As a result, the model remains competent simultaneously in text and sound — without compromise between the two areas.

For whom this is relevant

A unified multi-task model simplifies the development of voice and multimodal products. Instead of supporting multiple specialized systems, a team works with a single weight: easier to deploy, easier to update, easier to monitor. An update to one model automatically improves all five functions at once.

Potential application scenarios are broad: voice assistants with context understanding, real-time transcription and translation systems, multilingual voice platforms for call centers, audio content generation tools.

What this means

Audex continues the trend toward building multimodal foundational models with full task coverage. If before "audio + text" meant a combination of several specialized systems, NVIDIA offers a single weight, covering the full spectrum of audio tasks while preserving the text strength of the base model.

What can the Audex-30B neural network do?

Combines five types of audio tasks: understanding audio content, automatic speech recognition, synthesis, translation and additional sound processing — all in one model while preserving text abilities.