Authereon/Qwen3.6-35B-A3B-Fable-5-Distill-heretic

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

Authereon/Qwen3.6-35B-A3B-Fable-5-Distill-heretic is a 35.1 billion parameter decensored variant of the Qwen3.6-35B-A3B model, fine-tuned through distillation on Claude Fable 5 sessions. This model has undergone specific two-pass heretic directional ablation and subsequent coverage-weighted LoRA-DPO to suppress refusals and enhance injection defense. It is optimized for agentic deployments requiring robust tool-calling discipline and reduced safety filtering, while maintaining the original model's capabilities.

Loading preview...

Model Overview

Authereon/Qwen3.6-35B-A3B-Fable-5-Distill-heretic is a 35.1 billion parameter model derived from Qwen3.6-35B-A3B. It is a decensored version, fine-tuned using distillation from Claude Fable 5 sessions. The model has undergone significant modifications, including two-pass heretic directional ablation to suppress refusals and subsequent LoRA-DPO (Direct Preference Optimization) with coverage-weighted on-policy pairs to improve injection defense.

Key Capabilities & Features

  • Decensored Behavior: Deliberately reduced safety filtering, allowing it to answer content its stock sibling might refuse.
  • Enhanced Injection Defense: Achieves strong resistance against channel-impersonation compliance (1/30) and action laundering (0/30) through targeted DPO.
  • Robust Tool-Calling: Demonstrates 100% capture for tool-call discipline with served schemas, essential for agentic workflows.
  • Vision Capabilities: Supports vision tasks, including text OCR, chart reading, UI dialog, and shape recognition.
  • Claude Fable 5 Distillation: Inherits a bias towards code and tool-shaped structures due to training on coding-agent sessions.

Good For

  • Agentic Deployments: Ideal for applications requiring models to interact with tools and execute complex, multi-step instructions.
  • Use Cases Requiring Reduced Refusals: Suitable where a model needs to provide direct answers without excessive safety-induced refusals.
  • Research into Model Defenses: Provides a case study in applying ablation and DPO for injection resistance, though T7 exfiltration resistance remains a limitation requiring external filters.

Finetunes
1 models