Nemotron-Labs-Audex

"Unified Audio Intelligence Without Regressing on Text Intelligence" — a unified audio-text LLM that both understands and generates audio/speech (ASR, speech translation, TTS, general text-to-audio, speech-to-speech) in one decoder, with the headline claim that text reasoning does not regress. The 30B-A3B flagship is a hybrid Mamba-Transformer MoE on the Nemotron-Cascade-2 base (~3B active, 1M context, SFT+RL); the 2B is a dense SFT-only variant. Trained on 157.4B audio + 320.5B text tokens; uses external codecs (XCodec) rather than its own. NVIDIA OneWay Noncommercial License. Uploaded quietly July 6 with the arXiv report as the only announcement.