NVIDIA Unveils Audex-30B, an Integrated Audio-Text LLM

NVIDIA has unveiled Audex-30B-A3B, a new unified audio-text LLM built on Nemotron-Cascade-2-30B-A3B. The model uses a 30B-parameter MoE (Mixture-of-Experts) architecture, but achieves efficiency by using only 3B active parameters for actual computation.

Key Features and Capabilities:

• Multimodal Capability: Implemented by extending an audio encoder for speech and general audio input and a discrete audio token vocabulary for text/audio output.
• Wide Range of Audio Tasks: Supports audio understanding, speech recognition and translation, TTS (Text-to-Speech), audio generation, and Speech-to-Speech generation.
• Preserved Performance: Integrates audio capabilities while causing almost no degradation to the existing text-only model's reasoning, alignment, knowledge, long-context, and agentic capabilities.
• Technical Specifications: Supports a context length of up to 1M tokens, and supports both a Thinking mode using <think> tags and a regular Instruct mode.

Alongside this, NVIDIA is also providing a smaller model, Nemotron-Labs-Audex-2B.