LLaDA2.X: A Large-Scale Diffusion Language Model Series

Ant Group announced the LLaDA2.0 series, a milestone for discrete diffusion language models (dLLM). The series consists of LLaDA2.0-mini (16B) and LLaDA2.0-flash (100B), which applies a Mixture-of-Experts (MoE) architecture, marking the first case of scaling a diffusion model to the 100 billion parameter range.

Key features are as follows:

• 100B Parameter Scaling: LLaDA2.0-flash is the largest diffusion language model to date, and has demonstrated outstanding performance on code generation and complex instruction-following tasks.
• 2.1x Inference Acceleration: The LLaDA2.0-flash-CAP model, which leverages a parallel decoding mechanism, achieves speeds of up to 535 tokens/s, significantly outperforming existing AR (Autoregressive) models.
• Fully Open Source: Both the 16B and 100B model weights and training code have been made publicly available via Hugging Face.

For real-world deployment, custom inference engines based on dInfer and SGLang were built, supporting KV-Cache reuse and block-level parallel decoding to ensure practicality as a high-performance generative model.