Assiduous42/Qwen3.6-27B-Fable-Fusion-711-Heretic2

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Assiduous42/Qwen3.6-27B-Fable-Fusion-711-Heretic2 is a 27 billion parameter Qwen3.6-based language model with a 32K context length, specifically ablated to reduce creative-writing refusals. This model targets common refusal behaviors encountered during fiction generation, differing from standard AdvBench-style harm reduction. It achieves a significantly lower refusal rate of 1/100 on a held-out custom creative-writing prompt set, making it highly suitable for unconstrained creative text generation tasks.

Loading preview...

Qwen3.6-27B-Fable-Fusion-711-Heretic2: Creative Writing Refusal Ablation

This model is a specialized 27 billion parameter variant of the Qwen3.6 architecture, designed to address and significantly reduce refusals encountered during creative writing tasks. Unlike many "uncensored" models that primarily target AdvBench-style harmful content, this iteration focuses on eliminating refusals related to violence or adult themes within fictional narratives.

Key Differentiators & Capabilities

  • Targeted Refusal Reduction: The model was ablated using a custom probe set of ~350 creative-writing prompts where the base model previously refused, merged with AdvBench prompts for broader coverage.
  • Low Refusal Rate: Achieves a measured refusal rate of 1/100 on a held-out evaluation set of creative-writing prompts, a substantial improvement over the baseline of 17/100.
  • Minimal Capability Damage: The ablation process resulted in a very low KL divergence of 0.0173, indicating that the model's general capabilities remain largely intact.
  • Post-Training Ablation: The refusal ablation was performed as the final step, ensuring the reported metrics accurately reflect the model's behavior without subsequent fine-tuning reintroducing refusals.

Important Considerations

  • MTP Layer Removal: The Multi-Token Prediction (MTP) layer present in the base model was removed during the save_pretrained process, meaning speculative decoding via MTP is not available.
  • LoRA Application: Reproducing the ablation requires direct application of the LoRA math, as PeftModel.from_pretrained may silently fail to bind Heretic's key layout.

Ideal Use Cases

This model is particularly well-suited for developers and users requiring a large language model for:

  • Unconstrained Creative Writing: Generating fiction, roleplay scenarios, or narrative content that might involve themes often subject to refusal by more aligned models.
  • Storytelling & World-building: Creating detailed and unrestricted fictional universes without encountering content filters.
  • Dialogue Generation: Producing natural and uninhibited character interactions.