LLaDA MoE v2 Advances Diffusion Language Model Scaling

Fengqi Zhu, Shaoxuan Xu, Jingyang Ou, Zebin You, Yipeng Xing, Huabin Liu, Xiaolu Zhang, Jun Zhou, Zhenzhong Lan, Yankai Lin, Wayne Xin Zhao, Jianguo Li, Chongxuan Li, Ji-Rong Wen· August 5, 2026 View original

Key takeaways

  • MoE Diffusion Language Models (dLLMs) have distinct scaling behaviors from autoregressive models.
  • Optimal batch sizes and learning rates scale differently with compute for dLLMs.
  • LLaDA MoE v2 demonstrates strong performance with fewer pretraining tokens than leading AR models.
  • The research provides practical scaling laws and design principles for MoE dLLMs.

Who benefits

AI ResearchCloud ComputingSoftware DevelopmentContent CreationData Science

Summary

Researchers systematically characterized scaling behaviors for Mixture-of-Experts (MoE) Diffusion Language Models (dLLMs), identifying optimal hyperparameters and architectural choices. Their new model, LLaDA MoE v2, a 30B-A3B dLLM, demonstrates performance approaching leading autoregressive models like Qwen3 with fewer pretraining tokens.

Diffusion Language Models (dLLMs) offer an alternative to traditional autoregressive (AR) models, but their scaling properties, especially with Mixture-of-Experts (MoE) architectures, have been less understood. This research provides a systematic characterization of how various factors, including optimization hyperparameters, compute allocation, and architecture, influence the scaling of MoE dLLMs. Key findings reveal distinct scaling trends compared to AR models. For instance, optimal batch sizes grow faster, and learning rates decay more rapidly with increased compute. The study also suggests a slight data-side tilt in model-data allocation and indicates that larger scales increasingly favor larger expert pools while maintaining moderate expert granularity. Guided by these insights, the team developed LLaDA MoE v2, a 30B-A3B dLLM trained on 23.5 trillion tokens. This model achieves performance comparable to Qwen3 on several knowledge, reasoning, and coding benchmarks, despite using approximately 65% fewer pretraining tokens. After fine-tuning, it even surpasses SDAR Chat on multiple reasoning and coding tasks, establishing practical scaling laws and design principles for future MoE dLLMs.

Why it matters

This research provides crucial insights into efficiently scaling a promising alternative to autoregressive LLMs, potentially leading to more cost-effective and performant large language models for various applications.

How to implement this in your domain

  1. 1Review the identified scaling laws for MoE dLLMs to inform future model architecture and training decisions.
  2. 2Experiment with different optimization hyperparameters (batch size, learning rate) and compute allocations for dLLM training.
  3. 3Consider adopting MoE architectures with larger expert pools for very large-scale dLLM deployments.
  4. 4Evaluate the trade-offs between data-side and model-side compute allocation based on IsoFLOP analysis.
  5. 5Explore fine-tuning LLaDA MoE v2 or similar dLLMs for specific enterprise tasks to assess their performance against AR models.

Original post by Fengqi Zhu, Shaoxuan Xu, Jingyang Ou, Zebin You, Yipeng Xing, Huabin Liu, Xiaolu Zhang, Jun Zhou, Zhenzhong Lan, Yankai Lin, Wayne Xin Zhao, Jianguo Li, Chongxuan Li, Ji-Rong Wen

"arXiv:2608.03457v1 Announce Type: new Abstract: Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization h…"

View on X

Originally posted by Fengqi Zhu, Shaoxuan Xu, Jingyang Ou, Zebin You, Yipeng Xing, Huabin Liu, Xiaolu Zhang, Jun Zhou, Zhenzhong Lan, Yankai Lin, Wayne Xin Zhao, Jianguo Li, Chongxuan Li, Ji-Rong Wen on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses