Skip to content

Diffusal v1.0.0 — Initial public research release

Choose a tag to compare

View all tags
@wfzyx wfzyx released this 30 Jul 16:36
v1.0.0
2ae961d

Initial public research release

This release archives the Diffusal research program, including pre-registered experiment configurations, evaluation scripts, raw result summaries, and the current paper draft.

Current evidence

  • INT4 PTQ: matched MDLM diffusion language models show lower relative degradation than their autoregressive controls in the reported likelihood and generative-anchor evaluations.
  • Native ternary QAT: the three-seed 7M-scale replication supports the pre-registered no-extra-dLLM-tax criterion; it does not establish superior diffusion robustness.

Scope and limitations

Results are bounded by the evaluated models, hardware, and recipes. They are not a claim about 7B+ models, a 25B ternary dLLM, or calibrated PTQ. See ARXIV-PAPER.md and the experiment result files for methods, raw values, and caveats.

Citation and reuse

Citation metadata is available in CITATION.cff. Repository materials are available under CC BY 4.0.

License clarification: v1.0.0 was superseded by v1.0.1. Diffusal-authored code is MIT-licensed; paper and research prose are CC BY 4.0.

Assets 2
Loading