richardyoung/Qwable-9B-Claude-Fable-5-heretic
The richardyoung/Qwable-9B-Claude-Fable-5-heretic is a 9 billion parameter, decensored version of Empero's Qwable-9B-Claude-Fable-5 model, built using Heretic v1.4.0. This model is a full-parameter supervised fine-tune of Qwen3.5-9B, optimized for agentic coding and reasoning tasks with a 32768 token context length. It excels at imitating the reasoning and tool-use style of advanced assistants like Claude Fable 5 and GPT-5.5, making it suitable for complex coding and multi-turn agent tasks.
Loading preview...
Model Overview
This model, richardyoung/Qwable-9B-Claude-Fable-5-heretic, is a decensored variant of Empero's Qwable-9B-Claude-Fable-5, created using Heretic v1.4.0. It is a 9 billion parameter, full-parameter supervised fine-tune of the Qwen3.5-9B base model, which is natively multimodal with a hybrid attention stack and a large vocabulary. The fine-tuning focused exclusively on text, freezing the vision tower, and was performed on a curated dataset of agentic coding and reasoning traces from Claude Fable 5 and GPT-5.5 terminal agents.
Key Capabilities
- Agentic Coding & Reasoning: Specifically trained to imitate the reasoning and tool-use style of advanced assistants, excelling in long, multi-turn coding and agent tasks.
- Long Context Handling: Supports a native context length of 32,768 tokens, with training performed on full-length traces up to 76,800 tokens without truncation.
- Reasoning Model: Incorporates a
<think>block before generating final answers, providing transparent reasoning steps. - Decensored: This version has undergone a decensoring process, resulting in a lower refusal rate (50/100) compared to the original (97/100).
Good For
- Complex Code Generation: Generating correct and idiomatic code solutions, often incorporating current tooling and security-aware judgment.
- Terminal/Agentic Tasks: Solving problems that require multi-step agentic behavior and tool interaction.
- Instruction Following: Handling reasoning, instruction/format following, and calibration probes effectively.
- Research & Experimentation: Provided for community use in research and experimentation, particularly where a less restrictive model behavior is desired.