On iLLaDA, an 8B diffusion language model from researchers at Renmin University and ByteDance. Unlike standard autoregressive chat models, diffusion language models generate text through a different iterative process.

The report says iLLaDA keeps up with Qwen2.5 at the base level, though it falls behind after fine-tuning. That makes it an interesting research signal rather than a clear replacement for mainstream LLM architectures.

Alternative generation methods remain worth watching because they could eventually change latency, controllability, or training dynamics in language models.