Qwen3.8-27B Heretic

Derivative of Qwen/Qwen3.8-27B with refusal directions ablated by Heretic.

This is not a from-scratch trained model. All original weights and architecture are from the Qwen Team (Alibaba). Only directional-ablation parameters were applied.

Abliteration

• Tool: Heretic 1.4.0, 80 Optuna trials, 4-bit search on RTX 5090
• Saved trial: 30
• Refusals: 87/100 → 53/100
• KL divergence vs original: 0.0081
• Launch GGUF (local): Qwen3.8-27B-abliterated-Q6_K.gguf

License

Apache License 2.0, inherited from the base model.

Commercial use, modification, and redistribution are allowed with attribution. Keep this card, the Apache 2.0 LICENSE file, and the base_model metadata.

Attribution

• Base model: Qwen Team — Qwen3.8-27B
• Abliteration tool: p-e-w/heretic

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}