Because they shipped and are better only for specific applications. For everything else, they lag behind frontier models. They win in latency-sensitive niches (e.g., autocomplete, code infilling, subagent calls). Not frontier-level reasoning.

dLLMs are still promising. It's a counterintuitive hardware limitation. You should watch this: https://www.youtube.com/watch?v=r305-aQTaU0

They’re kinda dyslexic - like bad OCR