NVIDIA Releases Audex (Nemotron-Labs-Audex-30B-A3B): A Unified Audio-Text LLM That Preserves the Text Intelligence of Its Backbone

A unified audio-text language model was released that can both understand and generate audio and speech while maintaining the text performance of its underlying model. Most multimodal models experience a "text tax" where adding audio or vision capabilities causes text benchmark scores to drop, but this model was specifically designed to avoid that regression. The model uses a simple architecture that treats audio inputs and outputs similarly to text tokens, allowing it to run on standard language model software stacks. It was trained through multiple stages starting from a text-only foundation, followed by text-only reinforcement learning, which helped preserve strong text capabilities while adding audio abilities.