NVIDIA Releases Nemotron-Labs-Audex-30B-A3B: Unified Audio-Text MoE Model
NVIDIA
NVIDIA released Nemotron-Labs-Audex-30B-A3B on July 7 — a 30B Mixture-of-Experts model handling ASR, speech translation, TTS, text-to-audio generation, and speech-to-speech in a single decoder. Built on NVIDIA's Nemotron-Cascade-2-30B-A3B (hybrid Mamba-Transformer), 1M-token context, 3B active parameters. A smaller Audex-2B released simultaneously. Both on HuggingFace under noncommercial license.
Why it matters
Unified audio-language models that preserve strong text reasoning represent a meaningful architectural shift for voice agents and audio production tools. The 1M-token context and open noncommercial weights make Audex-30B accessible for academic and prototyping use at low inference cost (3B active params).
Importance: 3/5
First unified audio-text MoE from NVIDIA with 1M context; open noncommercial release covering the full audio task spectrum.