---

license: apache-2.0

base_model:

• Iloqt/Versi-StyleTune-31B

base_model_relation: quantized

tags:

- gguf

- gemma-4

- 31B

- merge

- mergekit

- reasoning

- creative writing

- roleplay

- conversational

not-for-all-audiences: true

---

Versi-StyleTune-31B (GGUF)

GGUF quants of Iloqt/Versi-StyleTune-head: a text-only Gemma 4 31B with Nimbz/Versipellis-31B as the base and the lm_head (output projection) grafted from Gryphe/Gemma-4-31B-StyleTune.

All credits go to Nimbz and Gryphe for the original models, I only committed the merge.

Variants

Standard quants:

• Q2_K, Q3_K_M, Q4_K_M, Q5_K_M, Q6_K, Q8_0 — body-only quantization.

hb16 variants (head + embeddings kept at BF16):

• Q4_K_M-hb16, Q5_K_M-hb16, Q6_K-hb16
• Preserves the grafted StyleTune lm_head and embed_tokens at full precision while quantizing the rest of the body to K-quant.

attn8-HB variants (Q8_0 attention + BF16 head + embeddings):

• Q4_K_M-attn8-HB, Q5_K_M-attn8-HB, Q6_K-attn8-HB
• Adds Q8_0 attention layers on top of the hb16 protection.

Notes

• Run with the Gemma 4 chat template; thinking off by default.
• Like all Gemma 4 models, benefits from repetition penalty or DRY to avoid token loops.