NVIDIA Audex unifies audio and text without sacrificing language ability

The buried detail is the training data composition: 157.4B audio tokens alongside 320.5B text tokens is a roughly 1:2 ratio, which is unusually audio-heavy for a model that still claims text parity. Most multimodal models that preserve text performance do so by keeping audio a minority signal, so the MoE architecture here is doing real work to prevent cross-modal interference, not just benefiting from a conservative data mix.

This sits in a cluster of work this site has tracked around making audio a first-class modality rather than a bolt-on. The Hugging Face and Cerebras piece from July 1 showed open-weight models reaching production-grade voice inference, but that integration kept audio and text as separate pipeline stages. Audex's single-decoder approach is architecturally more aggressive. Separately, the TiCodec paper from July 6 on streaming neural speech codecs addresses the inference-cost side of the same problem: once you have a unified audio-text model, efficient codec design determines whether it runs at acceptable latency. These two papers are effectively solving adjacent halves of the same deployment challenge.

Watch whether independent evaluators can reproduce the text benchmark parity on standard held-out suites like MMLU or HellaSwag without access to NVIDIA's training data, since the MoE routing strategy is the key variable and its behavior on purely text inputs is not yet externally verified.