NVIDIA Audex: one 30B MoE unifies text and audio reasoning

On July 6, NVIDIA released Nemotron-Labs-Audex-30B-A3B (arXiv 2607.05196), packing speech recognition, translation, TTS, audio generation, and speech-to-speech dialogue into a 30B MoE single Transformer decoder; through extended discrete audio vocabulary + projection to text embedding space + multi-stage SFT → Cascade RL → multi-domain on-policy distillation (157.4B audio + 320.5B text tokens), it reaches SOTA on six audio tasks with virtually no regression on text reasoning, reuses the vLLM/transformers 5.x inference stack, and is open-source but under NVIDIA's OneWay non-commercial license.