[!IMPORTANT] ⚠️ Warning: This model can produce narratives and RP that contain violent and graphic erotic content. Although this model hasn't been ablated yet using Heretic / SOMPOA, refusals can be reduced by using Gemma 4 26B NoThink chat template in combination with Jailbreaks or Supervised Reward Preferencing (SRP).

Goetia 26B-A4B v1.7

📜 Goetia 26B-A4B v1.7

⚗️ The Alchemical Crucible 🔥

This is a merge of pre-trained language models created using mergekit-exp, a custom fork of the official mergekit. This model was merged using the following (MoE-adapted) merge methods:

• della
• karcher
• nearswap
• multi_fusion

The following models were included in this merge:

• ConicCat/Gemma4-Writer-26BA4B
• Cyclone-Labs/Eclipsed-Phoenix-26B-A4B
• electroglyph/gemma4-26b-fiction-bf16
• google/gemma-4-26B-A4B-it
• Gryphe/Pantheon-Reasoning-26B-A4B-1.1-V2
• Hastagaras/G4-Jamet-26B-A4B-MK-1
• Naphula/Goetia-26B-A4B-v1.6
• Naphula/Promethean-Dawn-26B-A4B
• SubMaroon/Boulesis-v2.1-26B-A4B
• SubMaroon/Kitchoon-26B-A4B
• TheDrummer/Orion-26B-A4B-v1.1
• Vortex5/Shadow-Siren-26B-A4B

Thanks to the members of 26B-Suite for helping test and develop these models.

🙏 Thanks @mradermacher and other quantizers.

💾 GGUFs:
- mradermacher/Goetia-26B-A4B-v1.7-GGUF
- mradermacher/Goetia-26B-A4B-v1.7-i1-GGUF

⚙️ See this page for recommended settings.

🧙 Invocation Grimoires 🐙

🌋 Hellforged Parameters

mergekit_config_1_karcher.yml

architecture: Gemma4ForConditionalGeneration
merge_method: moe_karcher
# base_model: B:\26B\google--gemma-4-26B-A4B-it
models:
  - model: B:/26B/ConicCat--Gemma4-Writer-26BA4B
    parameters:
      weight: 1.0
  - model: B:/26B/Cyclone-Labs--Eclipsed-Phoenix-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/electroglyph--gemma4-26b-fiction-bf16
    parameters:
      weight: 1.0
  - model: B:/26B/Gryphe--Pantheon-Reasoning-26B-A4B-1.1-V2
    parameters:
      weight: 1.0
  - model: B:/26B/Hastagaras--G4-Jamet-26B-A4B-MK-1
    parameters:
      weight: 1.0
  - model: B:/26B/Naphula--Goetia-26B-A4B-v1.6
    parameters:
      weight: 1.0
  - model: B:/26B/Naphula--Promethean-Dawn-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/SubMaroon--Boulesis-v2.1-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/TheDrummer--Orion-26B-A4B-v1.1
    parameters:
      weight: 1.0
  - model: B:/26B/Vortex5--Shadow-Siren-26B-A4B
    parameters:
      weight: 1.0
parameters:
  max_iter: 20
  tol: 1.0e-9
  router_strategy: karcher  # Options: karcher, average, first, random_init
  blend_experts: true  # Blend corresponding experts (expert[0] + expert[0], etc.)
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
# chat_template: auto
trust_remote_code: true
name: 📜 Goetia 26B A4B v1.7A

mergekit_config_2_della.yml

architecture: Gemma4ForConditionalGeneration
base_model: B:/26B/google--gemma-4-26B-A4B-it
models:
  - model: B:/26B/ConicCat--Gemma4-Writer-26BA4B
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/Cyclone-Labs--Eclipsed-Phoenix-26B-A4B
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/electroglyph--gemma4-26b-fiction-bf16
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/Gryphe--Pantheon-Reasoning-26B-A4B-1.1-V2
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/Hastagaras--G4-Jamet-26B-A4B-MK-1
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/Naphula--Goetia-26B-A4B-v1.6
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/Naphula--Promethean-Dawn-26B-A4B
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/SubMaroon--Boulesis-v2.1-26B-A4B
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/TheDrummer--Orion-26B-A4B-v1.1
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
  - model: B:/26B/Vortex5--Shadow-Siren-26B-A4B
    parameters:
      weight:
        - filter: "lm_head"
          value: 0.1
        - filter: "attn"
          value: 0.1
        - filter: "mlp"
          value: 0.1
        - filter: "embed_tokens"
          value: 0.1
        - value: 0.1
      density: 0.9
      epsilon: 0.09
merge_method: moe_della
parameters:  
  lambda: 1.0
  # normalize: true # deprecated
  normalize_weights: false
  normalize_router: true
  int8_mask: false
  rescale: true
  router_strategy: della # average # random_init
  blend_experts: true
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
chat_template: auto
name: 📜 Goetia 26B A4B v1.7B

mergekit_config_3_multifusion.yml

architecture: Gemma4ForConditionalGeneration
merge_method: multi_fusion # v4
base_model: B:/26B/Goetia-26B-A4B-v1.7_karcher
models:
  - model: B:/26B/ConicCat--Gemma4-Writer-26BA4B
    parameters:
      weight: 1.0
  - model: B:/26B/Cyclone-Labs--Eclipsed-Phoenix-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/electroglyph--gemma4-26b-fiction-bf16
    parameters:
      weight: 1.0
  - model: B:/26B/Gryphe--Pantheon-Reasoning-26B-A4B-1.1-V2
    parameters:
      weight: 1.0
  - model: B:/26B/Hastagaras--G4-Jamet-26B-A4B-MK-1
    parameters:
      weight: 1.0
  - model: B:/26B/Naphula--Goetia-26B-A4B-v1.6
    parameters:
      weight: 1.0
  - model: B:/26B/Naphula--Promethean-Dawn-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/SubMaroon--Boulesis-v2.1-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/TheDrummer--Orion-26B-A4B-v1.1
    parameters:
      weight: 1.0
  - model: B:/26B/Vortex5--Shadow-Siren-26B-A4B
    parameters:
      weight: 1.0
  - model: B:/26B/Goetia-26B-A4B-v1.7_della
    parameters:
      weight: 2.0
parameters:
  importance_metric: "pareto_auto"    # Evaluates forward simulation loss across metric blends # 'pareto_auto', 'delta_mag', 'kl_div', 'cosine_sim', 'variance_sce', 'fisher_grad'
  mc_trials: 64                       # Deep 64-trial Monte Carlo continuous Dirichlet Pareto optimization
  power_of_two_scale: 8192            # High-resolution 8K sampling stride across tensors
  tukey_fence: 1.5                    # Strict outlier gate for high-definition novelty grafting
  anti_monopoly: 1.0                  # Unconstrained training retention (rewards genuine fine-tuning), lower values compress raw norm disparity across the donors (not recommended)
  sce_inversion_mode: true            # Inverts minority sign conflicts rather than deleting them
  is_moe: true                        # Set to false for dense models
  router_strategy: "metric"           # 'metric' (uses router_metric), 'average', 'first', 'base', 'random_init'
  router_metric: "kl_div"             # Evaluates exact categorical routing distribution on gate matrices # 'kl_div', 'pareto_auto', 'delta_mag', 'cosine_sim'
  blend_experts: true                 # Blend corresponding experts (expert[0] + expert[0], etc.)
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
# chat_template: auto
trust_remote_code: true
name: 📜 Goetia 26B A4B v1.7C

mergekit_config_4_nearswap.yml

architecture: Gemma4ForConditionalGeneration
merge_method: nearswap # v2
base_model: B:\26B\Goetia-26B-A4B-v1.7_multifusion
models:
  - model: B:\26B\SubMaroon--Kitchoon-26B-A4B  
    parameters:  
      t:  
        - filter: "layers.10.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.10.self_attn.v_proj"  
          value: 0.0001
        - filter: "layers.11.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.11.self_attn.v_proj"  
          value: 0.0001
        - filter: "layers.12.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.12.self_attn.v_proj"  
          value: 0.0001
        - filter: "layers.13.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.13.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.14.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.14.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.15.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.15.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.16.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.16.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.17.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.17.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.18.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.18.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.19.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.19.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.20.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.20.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.21.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.21.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.22.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.22.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.23.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.23.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.24.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.24.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.25.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.25.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.26.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.26.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.27.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.27.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.28.self_attn.o_proj"
          value: 0.0001
        - filter: "layers.28.self_attn.v_proj"
          value: 0.0001
        - filter: "layers.29.self_attn.o_proj"  
          value: 0.0001
        - filter: "layers.29.self_attn.v_proj"  
          value: 0.0001
        - filter: "lm_head"  
          value: 0.0001 
        - value: 0.0
parameters:
  router_strategy: "nearswap"
dtype: float32
out_dtype: bfloat16
tokenizer:
  source: union
chat_template: auto
trust_remote_code: true
name: 📜 Goetia 26B A4B v1.7D

🕵️ Pareto Telemetry Audit (1383)

---

🖼️ Ars Galleria