LLaDA2.2-flash Released

LLaDA2.2-flash, a new model in the LLaDA2 series, is a 100B-parameter (non-embedding) MoE architecture that adds agentic capabilities to a diffusion language model.

Key Features:

• 128K Context: Block Routing restricts MoE expert activation to the diffusion block level, improving efficiency for long-text processing
• Levenshtein Editing: Introduces DELETE/INSERT control tokens enabling sequence editing, deduplication, and insertion slot creation during parallel generation
• Agentic RL: Trained via L-EBPO (Levenshtein Editing ELBO-based block-level policy optimization) for multi-turn tool use and error correction

The primary goals are agentic applications such as multi-turn interaction, long-context tool use, and robust error correction.