Ant Group's inclusionAI Open-Sources LLaDA 2.2: World's First Large-Scale Agentic Diffusion Model with 128K Context

The autoregressive model's dominance in the AI agent arena is facing its first serious challenge. On July 28, Ant Group's inclusionAI team open-sourced LLaDA 2.2, a 100-billion-parameter MoE diffusion language model and the world's first large-scale Agentic diffusion model with native 128K context support.

The model can not only generate text in parallel but also self-modify and dynamically correct its outputs during generation. The team integrated the Levenshtein editing paradigm into the diffusion model's denoising process, supporting four atomic operations — KEEP, SUBSTITUTE, DELETE, and INSERT — giving the model the ability to revise its own outputs. On SWE-bench Verified, enabling Levenshtein editing alone yielded an absolute improvement from 35.8 to 44.4.

To address model collapse in long-horizon agent interactions, LLaDA 2.2 introduces L-EBPO, which frames multi-turn editing decisions as a reinforcement learning problem. This allows the model to autonomously decide when to remove redundant content or insert missing information based on environmental feedback.

On the engineering side, LLaDA 2.2 extends its context window from 8K to 128K through progressive long-context training. The team designed BlockRouting to select a fixed expert pool at the block level before performing token-level routing, significantly reducing HBM traffic and communication overhead.

In performance benchmarks, LLaDA 2.2-flash scored an average of 53.83 across seven agent benchmarks, compared to 55.74 for the leading autoregressive model Ling-2.6-flash — a gap of less than two points. On three interactive tasks the diffusion model actually outperformed its autoregressive counterpart.

Efficiency gains are even more pronounced: LLaDA 2.2-flash achieves an average BF16 throughput 1.64x that of Ling-2.6-flash across 11 workloads, with an additional 18.6% improvement under FP8 quantization.

The model weights, technical report, and code are all available open-source on GitHub and HuggingFace.