xtools-at/Qwen3.5-4B-Heretic-2xMerge-DavidAU

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

xtools-at/Qwen3.5-4B-Heretic-2xMerge-DavidAU is a 4.5 billion parameter language model based on Qwen3.5-4B, created by xtools-at through a weighted merge of two DavidAU fine-tunes. It combines Claude Opus-style reasoning distillation with the Deckard fine-tuning approach, both processed with HERETIC/abliteration to reduce refusal behavior. This experimental merge focuses on enhanced reasoning, general-purpose generation, creative writing, and roleplay, offering a compact model for diverse text generation tasks.

Loading preview...

Overview of Qwen3.5-4B-Heretic-2xMerge-DavidAU

This model is an experimental 4.5 billion parameter merge by xtools-at, built upon the Qwen3.5-4B base. It combines two distinct fine-tuned variants from DavidAU: Claude-4.6-OS and Deckard. Both source models were subjected to HERETIC/abliteration processing, aiming to reduce refusal behaviors while enhancing reasoning and output generation capabilities. The merge uses a linear weighted average to integrate these two approaches, resulting in a model designed for versatile text generation.

Key Capabilities

  • Enhanced Reasoning: Combines Claude Opus-style reasoning distillation with the Deckard fine-tuning methodology for improved logical deduction and analytical problem-solving.
  • Reduced Refusal Behavior: Both source models underwent HERETIC/abliteration processing, intended to mitigate common refusal patterns.
  • General-Purpose Generation: Capable of diverse text outputs, from instruction following to complex response composition.
  • Creative Writing & Roleplay: Optimized for generating stories, developing plots, character-driven interactions, and descriptive prose.
  • Multimodal Potential: Built on the Qwen3.5-4B architecture, it retains potential for image-and-text tasks, though specific multimodal functionality requires testing on the merged checkpoint.

Good For

  • Experimentation: Ideal for exploring the combined behaviors of different fine-tuning strategies in a compact model.
  • Creative Content Generation: Storytelling, scriptwriting, and generating engaging dialogue.
  • Role-Playing Scenarios: Creating dynamic character interactions and immersive conversational experiences.
  • Reasoning Tasks: Tackling multi-step questions and logical challenges where a less restrictive output is desired.
  • Instruction Following: Generating varied responses based on given prompts and instructions.

Finetunes
2 models