iLLaDA diffusion model matches autoregressive AI at 8-billion-parameter scale

iLLaDA, an eight-billion-parameter diffusion language model, generates text by repeatedly refining a masked passage rather than predicting words left to right. Released on June 25, 2026, with weights and code on arXiv, it is an improved successor to the earlier LLaDA model and was trained entirely using the diffusion approach. The model performs competitively with a similarly sized conventional autoregressive model across general knowledge, math, and coding benchmarks — a first for diffusion-based language models at this scale. Researchers argue the architecture has inherent advantages for long-range planning and bidirectional reasoning, though the comparison holds only when both models are matched on compute and training data. The result suggests diffusion language models are a second viable architectural path alongside the autoregressive approach that has dominated the AI chatbot era.