---

base_model: Vortex5/Gemma-4-26B-A4B-Animus-V14.1-FFT-heretic

language:

- en

- ru

library_name: transformers

tags:

- mergekit

- task_arithmetic

- gemma

- gemma4

- moe

- roleplay

- russian

- english

- gguf

- quantized

license: gemma

---

Kanimus-26B-A4B-FFT-heretic (v4 QK-LoRA Merge)

An experimental Dark Roleplay model created by manually merging QK projections via task-arithmetic from a Claude Opus distillation, followed by the integration of a stylistic LoRA on top of the Animus-heretic base.

Architecture and Methodology

The model is built upon Gemma 4 26B A4B. The merge arithmetic was executed in FP32, with the final weights saved in bfloat16.

Step 1: QK Injection (Task Arithmetic)

A manual task-arithmetic injection was applied strictly to the Attention Query and Key projections. Formula used:

W_final = W_animus + α * (W_opus - W_unsloth)

Reference models:

• Parent Opus: unsloth/gemma-4-26B-A4B-it
• Target Opus: TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill-v2

Separate α coefficients were applied for a controlled attention shift (respecting QK-norm):

• α_sliding_q = 0.50
• α_sliding_k = 0.50
• α_global_q = 0.50
• α_global_k = 0.25 (halved to mitigate KV-injection collapse on long contexts)

Post-merge Invariant: All non-QK tensors (including MoE router, shared MLP, embeddings, and norms) remain bitwise equal to the original Animus-heretic base. This was verified via torch.equal during assembly.

Step 2: LoRA Bake

The stylistic LoRA SubMaroon/Dark-Goetia-26B-A4B-LoRA-v4 was baked directly into the model weights.

• Method: set_scale("default", 0.20) -> merge_and_unload(safe_merge=True)
• Effective weight: 0.40 * BA (accounting for alpha/r = 2).

Usage and Recommendations

The model is designed for English and Russian Roleplay. It maintains scene logic, exhibits high NPC initiative, and operates within a dark, atmospheric writing style.

Recommended Sampling:

• Temperature: 0.7 - 0.8
• min_p: 0.05 - 0.1
• Repetition penalty: 1.05 - 1.10

Available Versions

• BF16: Full-precision version (~53 GB).
• GGUF Q4_K_M.

License

Distributed under the Gemma license. Usage is subject to the terms of the original Google models and the respective distillates.

Acknowledgements

• Darkhn for the Animus V14.1 base model.
• Vortex5 for the Animus V14.1 Heretic model.
• TeichAI for the Opus Distill v2.