NVIDIA has released Audex (Nemotron-Labs-Audex-30B-A3B), a 30B MoE language model with 3B active parameters that unifies audio and text processing while preserving text reasoning capabilities. Built on the text-only Nemotron-Cascade-2-30B-A3B backbone, Audex projects audio inputs into the text space, avoiding the performance degradation seen in other unified audio models.