MuXodious/Qwen3.8-27B-absolute-heresy

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

MuXodious/Qwen3.8-27B-absolute-heresy is a fine-tuned version of the Qwen3.8-27B model, developed by MuXodious using P-E-W's Heretic v1.4.0 ablation engine with Self-Organizing Maps & Magnitude-Preserving Orthogonal Ablation. This 27 billion parameter model is specifically optimized to reduce refusals, achieving an "Absolute Heresy" index with near-zero overt and secondary refusals while maintaining minimal model damage. It excels in agentic coding, long-horizon office work, and multimodal intelligence tasks, offering native vision-language capabilities and an extensible context length up to 1,000,000 tokens.

Loading preview...

Overview

MuXodious/Qwen3.8-27B-absolute-heresy is a fine-tuned variant of the Qwen3.8-27B model, developed by MuXodious. This model leverages P-E-W's Heretic ablation engine, specifically with Self-Organizing Maps & Magnitude-Preserving Orthogonal Ablation, to significantly reduce model refusals. It achieves an "Absolute Heresy" index, indicating near-zero overt and secondary refusals with minimal impact on core model performance.

Key Capabilities

  • Refusal Reduction: Engineered to minimize refusals, making it highly compliant for various tasks.
  • Agentic Performance: Demonstrates strong capabilities in agentic coding (e.g., SWE-bench Pro: 61.7, DeepSWE 1.1: 42.2, QwenSWEBench: 79.0) and long-horizon office work (CoWorkBench: 70.7).
  • Multimodal Intelligence: Features native vision-language understanding, supporting image and video inputs, and excels in tasks like computer use (OSWorld-Verified: 84.3) and browser use (WebArena-Verified: 64.8).
  • Flexible Thinking Control: Supports adjustable reasoning depth (reasoning_effort) and preserves reasoning context across turns (preserve_thinking).
  • Extended Context Length: Natively supports 262,144 tokens, extensible up to 1,000,000 tokens using YaRN scaling.

Good For

  • Applications requiring highly compliant and non-refusing AI responses.
  • Complex agentic workflows, including coding, software engineering, and office automation.
  • Multimodal tasks involving image and video analysis, such as visual math problem-solving (MathVision With CI: 94.6) and general visual reasoning (BabyVision With CI: 85.6).
  • Developers seeking a powerful Qwen3.8 base model with enhanced control over reasoning and refusal behavior.

Finetunes
7 models