Ant Group’s InclusionAI team released LLaDA2.2 this month, the third major version in a lineup of diffusion language models that now spans seven public checkpoints running from 16 billion to 100 billion parameters. The family’s flagship, LLaDA2.0-flash, is a Mixture-of-Experts model that InclusionAI calls the first diffusion model to reach the 100 billion parameter mark. That scaling claim is notable because diffusion architectures have largely stayed at research scale while autoregressive labs pushed model sizes into the hundreds of billions.
Standard large language models generate text autoregressively: they predict one token at a time, left to right, with each new token conditioned on everything written before it. Diffusion language models work differently. They start from a fully masked or noisy sequence and refine many positions at once across a series of denoising steps, letting a block of tokens update in parallel instead of waiting on a strict left-to-right chain. That structural difference is why the architecture draws attention for latency, since fewer sequential steps can translate into faster generation for a given output length. It also suits agent workflows, where a system often needs to revise or patch part of an already-generated plan or piece of code rather than only extend it, an operation that maps more naturally onto denoising than onto autoregressive continuation.
InclusionAI has shipped three major versions in nine months. LLaDA2.0 arrived in November 2025 as the base Mixture-of-Experts family. LLaDA2.1 followed in February 2026 with what the team calls token editing, aimed at speeding up the diffusion process itself. LLaDA2.2, released this month, adds a technique InclusionAI calls Levenshtein Editing, which the team ties directly to agentic use: modifying prior output rather than regenerating it from scratch.
On performance, InclusionAI reports a 2.1x inference speedup from a parallel decoding mechanism, with a variant called LLaDA2.0-flash-CAP reaching up to 535 tokens per second. The gain comes from a custom inference stack built on dInfer and SGLang, featuring reused KV-cache and decoding carried out in parallel blocks. The repository does not include independent benchmark comparisons against similarly sized autoregressive models. The figures published so far are InclusionAI’s own.
All seven checkpoints, from the 16 billion parameter LLaDA2.0-mini through the 100 billion parameter flash variants, are published on Hugging Face along with training code, under an Apache 2.0 license that permits commercial use. That combination of full weights, training code, and a permissive license is what will let outside developers test InclusionAI’s speed claims for themselves.
Teams building latency-sensitive agent products now have a fourth architecture family to benchmark against their current autoregressive stack, one built specifically around editing outputs rather than only generating them.
InclusionAI, the Ant Group team behind the LLaDA2.X repository on GitHub, published the LLaDA2.2 update in July 2026.
