Gryphe, who publishes as Gryphe Padar, develops language-model merging tools and creative-writing models. Their work includes MythoMax, BlockMerge Gradient and MergeMonster, along with later fine-tunes and synthetic datasets.[1][2][3]
Start with MergeMonster →
Links
Hugging Face
GitHub Contributions
Combine model weights through a recorded recipe
MythoMax combines MythoLogic and Huginn using an experimental merging approach. Its model card explains that different tensors receive different interpolated weights and links to the configuration. This makes the recipe inspectable, while leaving evaluation of the resulting writing behavior to tasks and examples rather than a claim that merging always improves a model.[2]
Optimize a merge against example phrases
MergeMonster evaluates candidate merges using lists of desirable and undesirable phrases. It retains changes according to the configured objective, working through model layers and possible interpolation methods. That objective can shape particular output tendencies, but improving a phrase-based score does not establish broad reasoning or writing quality.[3]
Publish experiments beyond the original merge
Gryphe’s current publisher profile lists Pantheon and StyleTune models and describes the releases as personal fine-tunes. It also publishes synthetic datasets for writing and roleplay. These give readers a way to follow the creator’s later experiments alongside the older MythoMax recipe.[1]
Tools, writing & talks
Tools
Start here
MergeMonstergithub.comA configurable search over model merges guided by desired and undesired phrases.
Read the phrase-based objective and configuration before attempting a merge. Python and compatible checkpoints are required. The README lists substantial GPU-memory requirements; its smallest stated mode still needs room for a complete model.
Developer[3]- MythoMax-L2-13Bhuggingface.co
A documented model merge combining MythoLogic and Huginn.
Model creator[2] - BlockMerge Gradientgithub.com
Model-merging code and configurations that expose per-layer mixing choices.
Developer[4]