NVIDIA Releases Audex (Nemotron-Labs-Audex-30B-A3B): A Unified Audio-Text LLM That Preserves the Text Intelligence of Its Backbone

NVIDIA has introduced Audex (Nemotron-Labs-Audex-30B-A3B), a groundbreaking multimodal language model that integrates audio and text processing into a single unified architecture. This new model represents a significant advancement in AI's ability to handle diverse audio and language tasks simultaneously while maintaining the linguistic intelligence of its predecessor. The 30-billion parameter mixture-of-experts (MoE) model builds on NVIDIA's Nemotron-Cascade-2 backbone, establishing a new standard for comprehensive audio-text understanding.

Audex consolidates five primary capabilities into one cohesive model: audio understanding, automatic speech recognition (ASR), speech translation, text-to-speech (TTS) synthesis, and audio generation. The model employs a mixture-of-experts architecture to optimize computational efficiency while processing complex multimodal inputs. A critical technical achievement is that Audex preserves the text-based reasoning capabilities of its Nemotron-Cascade-2 foundation with only marginal performance regression, meaning users don't sacrifice linguistic intelligence when gaining audio processing abilities.

• Unified AI Pipeline: Consolidates previously separate models into one architecture, reducing complexity in deployment and resource requirements
• Preserved Performance: Maintains strong text understanding despite expanded audio capabilities, eliminating typical trade-offs in multimodal models
• Enhanced User Experience: Enables more natural human-computer interaction through simultaneous audio and text processing
• Development Efficiency: Reduces training and inference costs by combining multiple specialized models into a single MoE framework
• Competitive Advantage: Positions NVIDIA as a leader in practical multimodal AI solutions for enterprise applications

Audex addresses a fundamental challenge in modern AI development: creating models that excel across multiple modalities without sacrificing performance in any single area. As businesses increasingly demand conversational AI systems that can handle voice commands, transcription, translation, and speech synthesis, having a unified model reduces infrastructure complexity and operational costs. NVIDIA's ability to maintain text intelligence while expanding into audio processing demonstrates sophisticated engineering that could influence how enterprises design their next-generation AI systems. This release signals NVIDIA's commitment to delivering practical, efficient multimodal solutions that advance the boundaries of what unified language models can accomplish.