DMax
Summary
DMax is the diffusion-language-model method family introduced by DMax: Aggressive Parallel Decoding for dLLMs. It combines On-Policy Uniform Training with Soft Parallel Decoding so a masked diffusion language model can revise its own intermediate predictions instead of committing every mask-to-token decision irreversibly.
Role In The Wiki
Use this page as the object card for the DMax method and released model family. The source page carries the detailed evidence, limitations, and agenda mapping.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in dmax-2026 rather than duplicating verdict rows here. At the entity level, DMax is relevant as an adjacent dynamic-compute and generation mechanism: it shows how a diffusion language model can increase decoding parallelism while preserving quality by keeping intermediate states revisable. It is not direct evidence for numeric time-series modeling, native multivariate state, event streams, or action-conditioned rollouts.
Evidence
Official Artifacts
- Preprint: arXiv 2604.08302
- Official code: czg1225/DMax — GitHub API reported Apache-2.0 license and 126 stars at ingest time.
- Official general model: Zigeng/DMax-16B — Hugging Face API reported
license:apache-2.0,base_model:inclusionAI/LLaDA2.0-mini, andcustom_codetags at ingest time. - Official math model: Zigeng/DMax-Math-16B — specialized for math and reasoning.
- Official code model: Zigeng/DMax-Coder-16B — specialized for code generation.
- Official math training data: DMax LLaDA-2.0 Mini Math Trajectories.
- Official code training data: DMax LLaDA-2.0 Mini Code Trajectories.