richardyoung/Qwable-9B-Claude-Fable-5-heretic

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The richardyoung/Qwable-9B-Claude-Fable-5-heretic is a 9 billion parameter, decensored version of Empero AI's Qwable-9B-Claude-Fable-5, built using Heretic v1.4.0. This model is a supervised fine-tune of Qwen3.5-9B, optimized for agentic coding and reasoning tasks by imitating Claude Fable 5 and GPT-5.5 terminal agent outputs. It features a 32768 token context length and excels at generating structured reasoning blocks before providing solutions.

Loading preview...

Overview

This model, richardyoung/Qwable-9B-Claude-Fable-5-heretic, is a decensored variant of Empero AI's Qwable-9B-Claude-Fable-5, created with Heretic v1.4.0. It is a 9 billion parameter, full-parameter supervised fine-tune of Qwen/Qwen3.5-9B, a natively multimodal model with a hybrid attention stack and ~152k vocabulary. The fine-tuning focused exclusively on text, freezing the vision tower.

Key Capabilities & Training

  • Agentic Coding & Reasoning: Fine-tuned on a curated mix of agentic coding and reasoning traces from Claude Fable 5 and GPT-5.5 terminal agents. It learns to imitate their reasoning and tool-use style.
  • Long Context: Supports a maximum sequence length of 76,800 tokens, with no truncation during training.
  • Reasoning Structure: Every response begins with a <think>...</think> block, providing a transparent reasoning process before the final answer.
  • Decensored: This version specifically aims to reduce refusals compared to the original model, achieving 50/100 refusals versus 97/100 in the original.
  • Reproducible: The model's creation process is documented for reproducibility.

Performance & Limitations

  • Strong Coding: Qualitative reviews indicate strong performance in coding and terminal/agentic prompts, providing correct and idiomatic solutions.
  • Text-Only Fine-tune: While the base model is multimodal, only the text path was trained; vision capabilities are inherited but not tuned or tested.
  • Domain-Focused: Its strengths are concentrated in coding and agentic tasks. For general knowledge, verification is recommended.
  • Inherited Style: Reflects the style and limitations of its base (Qwen3.5-9B) and teachers (Claude Fable 5, GPT-5.5), with no additional safety tuning beyond the base model.

Finetunes
1 models