the diffusion versus autoregressive debate finally has a clean data point, and it points to a much narrower claim than the hype

For about a year the diffusion versus autoregressive argument has mostly run on vibes. One camp says next token prediction is a dead end and parallel denoising is the future, the other says diffusion never survives contact with real tasks. Almost nobody was putting the two side by side on the same evals from the same lab. The lab behind it just did exactly that, with a model they pushed out only hours ago, and the honest part is what they compared against. They benchmarked their diffusion model, the one they call LLaDA2.2, against their own autoregressive model of similar size. So this is not a vendor picking a weak outside competitor. It is one lab grading its two bets against each other. The result is more useful for being unflattering. On general knowledge and most coding evals the diffusion model trails its autoregressive sibling. It only pulls ahead on a handful of interactive agent benchmarks, winning tau2 bench at 80.33 over 76.36, and MCP Atlas at 46.21 over 41.12, the multi turn tool calling style tasks. The SWE bench gap looks worse than it really is because the two runs used different scaffolds, so I would not read that one literally. Where it clearly wins is speed. Roughly 1.6x average decoding throughput over the autoregressive sibling in BF16 with speculative decoding on, and up to about 2.3x on the agent workloads. It gets there by keeping, substituting, deleting and inserting tokens inside a block instead of committing to each token once and never revisiting it. The weights are open under Apache 2.0, which is the only reason any of this is checkable, but it is a 205.8 GB, 100B class download with no llama.cpp support and only a coming soon serving story, so almost nobody runs it locally, and structured output is weaker than the autoregressive baseline by the authors' own admission. So I do not read this as diffusion wins. The narrower claim it actually supports is that for agent loops where you pay decode latency every turn, a diffusion backbone can reach a comparable quality band noticeably faster. That is a niche, not a paradigm shift, and the paradigm shift was never the honest question. What we finally have is a way to measure the tradeoff instead of arguing about it.