LLaDA2.2 just dropped, a 100B class diffusion model for agents that edits its own tokens while it decodes

Article automatically generated from technical news.

This one only landed a few hours ago, and the first thing I went looking for was the serving story rather than the benchmarks. It is a 100B class Mixture of Experts diffusion model aimed squarely at agent work, and the genuinely novel part is that it revises its own draft while it decodes. The four moves it makes inside a block are keep, delete, insert, and substitute, so it can pull redundant spans, open new editable slots, and swap tokens instead of freezing everything in one left to right