Nvidia launches Audex, a unified audio-text AI model with 30B parameters

Nvidia released a paper in June introducing Audex (Nemotron-Labs-Audex-30B-A3B), a unified audio-text language model. It has 30 billion total parameters and 3 billion active parameters, designed to understand and generate audio and speech while keeping core text smarts. The backbone model is Nemotron-Cascade-2-30B-A3B, with 52 layers of mixed Mamba-Transformer structure and 128 routed experts, activating 6 per step. The goal is to avoid text ability drop after adding multimodal features. For audio, it uses AF-Whisper to process 16kHz input and X-Codec2 for speech output at 50 tokens per second. In tests, Audex scored 86.4 on MMLU-Redux, slightly above the backbone's 86.3, but saw small drops on MMLU-Pro and GPQA-Diamond.