NVIDIA’s Audex: The End of the "Audio Tax"?

NVIDIA's Nemotron-Labs-Audex-30B-A3B release notes reveal a massive leap for unified models. Most multimodal attempts to add audio capabilities result in a "text tax"—where the model gains hearing and speaking abilities at the direct expense of its core reasoning and intelligence.

But Audex is different. NVIDIA used a MoE (Mixture-of-Experts) architecture that projects audio into the text embedding space instead of just stacking a separate "talker" module on top. The result? It actually maintains its text intelligence. In fact, looking at the benchmarks, it actually slightly outperformed its text-only backbone on some reasoning tasks. That is wild.

What really caught my eye is how it handles general audio. While most models focus strictly on speech-to-speech, Audex uses X-Codec to generate general audio sounds, making it much more versatile. It even managed to beat Qwen3-Omni on several coding and instruction benchmarks despite being a unified model.

If this holds up in real-world latency tests, we might be looking at the new standard for open-source omni-models. It’s not just a chatbot that can talk; it’s a reasoning engine that actually understands the world through sound.

What do you guys think? Is the era of specialized "thinking" vs "talking" models coming to an end?