LLaDA2.2-flash

model
Ant Group

The first agentic diffusion language model in Ant's LLaDA2 line: Levenshtein-style editing with DELETE/INSERT tokens, 128K context via Block Routing, and L-EBPO agentic RL. SWE-bench Verified 49.28 under a Claude Code scaffold at ~519 tokens/sec — roughly 1.7× its autoregressive sibling's speed.

HuggingFaceHuggingFace (LLaDA2.2-mini)Technical Report (PDF)

Model Details

Context window 128,000

Variants

Name Parameters Notes
LLaDA2.2-flash 100B Jul 22 2026; the file's anchor
LLaDA2.2-mini 16B Sep 5 2026; lightweight variant on the LLaDA2.0-mini architecture with the same Levenshtein editing, Block Routing 128K context, and L-EBPO agentic RL; 256 experts, 8 active; Apache 2.0
open-weightagentsefficiencyresearch

Related