https://introspective-diffusion.github.io/
Article
-
New approach converts autoregressive LLMs (Qwen) into diffusion models
-
I-DLM-8B and I-DLM-32B released on HuggingFace (April 2025)
-
Vastly outperforms native diffusion models, competitive with base autoregressive model
Discussion
-
Commenters excited: key upside is parallel token generation enabling much faster inference
-
Questions on whether diffusion enables iterative self-correction during generation
-
Uncertainty about whether this is practically usable as a drop-in faster Qwen replacement
Discuss on HN