What's this about?
Token Time Continuous Diffusion is a new diffusion-based language model: it operates in continuous space, deterministically transforming Gaussian noise into a finished token canvas, with a novel per-token notion of time (HF 7↑).
What exactly is a "diffusion language model"?
Mainstream LLMs write left-to-right, one token at a time (autoregressive). Diffusion models — the tech behind image generators — start with pure noise and iteratively refine the whole canvas at once. Applied to text, that promises parallel generation (speed) and global revision (coherence), and labs are actively exploring it (see DeepMind's DiffusionGemma). TTCD's twist is giving each token its own "clock" during the denoising process.
Why does it matter?
Diffusion LMs are the most credible challenger to the autoregressive monopoly; per-token time is the kind of architectural idea that, if it improves quality-vs-speed trade-offs, propagates fast.