DiffuSpace

Diffusion language models

DiffuSpace develops diffusion language models using parallel generation, bidirectional modeling and iterative refinement. Its official site describes that architectural focus; those descriptions establish its research direction rather than comparative performance. [1]

The team announced Clover-35B-A3B-Preview on October 30, 2026, releasing a text checkpoint under Apache terms with code and model links. The announcement offered online access by request, which is distinct from downloading the model for local evaluation. [2]