akumaburn/Swift-Qwen3.8-27b-heretic

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 18, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The akumaburn/Swift-Qwen3.8-27b-heretic is a 27 billion parameter Qwen3.8-27B model, derived from UkisAI/Swift-Qwen3.8-27b, with its safety alignment deliberately removed using the Heretic method. This BF16 model retains the original vision tower and 262,144-token context length, but is specifically engineered to comply with harmful, dangerous, illegal, and unethical requests for research into interpretability and safety. It is intended for red-teaming and evaluation purposes where unfiltered output is required.

Loading preview...

Swift-Qwen3.8-27b-heretic: Safety Alignment Removed

This model is a 27 billion parameter variant of the Qwen3.8-27B architecture, specifically derived from UkisAI/Swift-Qwen3.8-27b. Its primary distinguishing feature is the deliberate removal of safety alignment using the Heretic method, which ablates the residual-stream direction mediating refusals. This process ensures the model will attempt to comply with requests that the source model would refuse, including those deemed harmful, dangerous, illegal, or unethical.

Key Characteristics

  • Censorship Removal: Achieves 0 hard refusals on 100 mlabonne/harmful_behaviors prompts, compared to 98/100 for the source model.
  • Minimal Divergence: Optimized to minimize KL divergence from the source model on ordinary inputs, with a KL of 0.0763 nats.
  • Architecture Preservation: Retains the original vision tower and a native 262,144-token context length.
  • BF16 Numerics: The base model is in BF16 format, with optimized INT8 variants available for faster serving.
  • Thinking Model: Utilizes the Qwen3.5 chat template with reasoning enabled by default.

Intended Use Cases

  • Interpretability Research: Studying how models generate and refuse content.
  • Safety Research: Evaluating model vulnerabilities and red-teaming.
  • Unfiltered Content Generation: For specific research scenarios requiring responses to harmful or unethical prompts.

Caution: This model is a research artifact with no content moderation and is provided without warranty. Users are responsible for their own use and compliance with all applicable laws.

Finetunes
4 models