World’s First Agentic Diffusion Model Arrives: Correcting Errors While Acting, 128K Context Matches Autoregressive Models

InclusionAI, backed by Ant, unveiled LLaDA2.2, a 100B-scale MoE diffusion language model with native 128K context support.

The system combines Levenshtein editing, reinforcement learning from environment feedback, and BlockRouting so the model can add, remove, and correct tokens during generation.

On several agent benchmarks and throughput tests, it reportedly matches or even beats leading autoregressive models in some settings.

The result suggests diffusion models are becoming viable for long-horizon agent tasks, opening the door to a two-track architecture alongside autoregressive models.