Done by request, enjoy.

Barcenas-StyleTune-31B-Fable

A text-only Gemma 4 31B model: Danielbrdz/Barcenas-31b-Fable as the base, with the output projection (lm_head) grafted from Gryphe/Gemma-4-31B-StyleTune.

Recipe

• Base: Barcenas-31b-Fable, text layer extracted (vision tower removed).
• Output head: StyleTune lm_head.weight, with tie_word_embeddings set to false.
• No layer-level merging — Barcenas weights are otherwise unchanged.

Notes

• Text-only (vision/audio removed).
• Run with the Gemma 4 chat template.
• Like all Gemma 4 models, benefits from repetition penalty or DRY to avoid token loops.