LLaDA2.2-flash (128K)

LLaDA2.2-flash — an agent-oriented diffusion LM with a 128K context window, introducing Levenshtein editing (DELETE / INSERT control tokens) for long-context tool use, multi-turn interaction and robust error correction.

Overview

LLaDA2.2 is the agent-oriented generation of the LLaDA2 diffusion language model series, taking the family’s first step into agentic applications. By introducing Levenshtein editing — with DELETE and INSERT control tokens — to diffusion language modeling, LLaDA2.2-flash supports long-context tool use, multi-turn interaction, and robust error correction.

The model natively operates over a 128K context window and is positioned for agentic workloads: on τ²-Bench it scores 80.33 and on SWE-bench Verified 49.28 (Claude Code scaffold), with a throughput advantage over dense AR baselines thanks to its sparse MoE diffusion design.

LLaDA2.2 replaces the pure “remask” loop with Levenshtein editing: the model can delete and insert tokens during decoding, which is essential for agents that must revise earlier steps.

Levenshtein editing with KEEP / DELETE / INSERT control tokens, a training cluster, a sandbox cluster and a black-box agentic runtime over 300B / 64K long-context data. Source: model card.

Model Series

Model Notes
LLaDA2.2-flash Flagship agent-oriented dLLM (128K context)
LLaDA2.2-mini Lightweight variant

Highlights

SWE-bench Verified / Pro / Multilingual and throughput against an AR baseline (Claude Code scaffold, 128K context, average of five runs). Source: model card.

Resources