InclusionAI Releases LLaDA2.2-flash, an Agent-Oriented Diffusion Language Model
inclusionAI has released LLaDA2.2-flash, an agent-oriented diffusion language model featuring Levenshtein Editing and Block Routing Mixture-of-Experts (MoE). It is designed to support efficient 128K long-context agentic workloads and robust error correction. This release represents a significant step forward in non-autoregressive LLM research by enabling diffusion models to handle complex multi-turn interactions and tool use. The integration of Levenshtein Editing allows the model to dynamically edit sequence structures and correct errors during parallel generation. LLaDA2.2-flash features 100 billion non-embedding parameters, a 128K token context window, and uses DELETE and INSERT control tokens for sequence editing. It is trained using a new reinforcement learning framework called Levenshtein Editing ELBO-based Block-level Policy Optimization (L-EBPO).
## BACKGROUND
Diffusion language models adapt diffusion processes, originally used for image generation, to generate text. While traditional Mixture-of-Experts (MoE) models route each token individually, Block Routing MoE routes experts once per decoded block to prevent memory overhead in diffusion-based architectures.