OPEN SOURCE
Az Ant Group kiadta a LLaDA2.1 szerkeszthető diffúziós nyelvmodelleket
Ant Group released LLaDA2.1, a discrete diffusion language model that enables dynamic error correction during generation through a novel Token-to-Token editing mechanism. The system incorporates the first large-scale reinforcement learning framework for diffusion language models, using ELBO-based Block-level Policy Optimization to improve reasoning and instruction-following capabilities across 33 benchmarks. The model operates in two primary modes to balance speed and accuracy.
- LLaDA2.1-Mini (16B parameters) reaches peak speeds exceeding 1,500 tokens per second
- Speedy Mode uses aggressive confidence thresholds for rapid drafting
- Quality Mode maintains conservative thresholds for superior benchmark performance
- Uses a Token-to-Token editing mechanism for dynamic error correction