According to the arXiv abstract, Diffuplex is a full-duplex spoken language model based on discrete diffusion. The authors claim it achieves an unprecedented balance between linguistic intelligence and response efficiency, suggesting that the model can handle conversational turns without sacrificing comprehension or speed.

The abstract mentions a 'thinking mode' in which Diffuplex delivers a 13.0 result. However, the text truncates at that point, so the specific benchmark or unit is not provided in the available version.

As a single-source announcement, the claims are preliminary and have not been independently verified. The key contribution appears to be the application of discrete diffusion to full-duplex speech, a direction that could matter for real-time spoken dialogue systems.