nurdich/Qwen3.8-9B-Distill-uncensored-heretic

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

nurdich/Qwen3.8-9B-Distill-uncensored-heretic is a 9 billion parameter language model based on the Qwen3.5 architecture, developed by nurdich. This model is a censorship-ablated version of Empero's Qwen3.8-9B-Distill, specifically designed to reduce refusal rates while maintaining a low deviation from the base model's original behavior. It excels in reasoning tasks, particularly mathematics and code, by inheriting distilled chain-of-thought capabilities from a frontier-scale teacher model, making it suitable for complex problem-solving on a single GPU.

Loading preview...

Model Overview

nurdich/Qwen3.8-9B-Distill-uncensored-heretic is a 9 billion parameter model derived from Empero's Qwen3.8-9B-Distill, which itself is a full-parameter distillation of a larger Qwen3.8 2.4T A95B teacher model. This specific variant has undergone a censorship ablation process using the heretic method to significantly reduce refusal rates, achieving a refusal rate of 6/100, comparable to the original Qwen3.5-9B base, but with a 2.6x lower KL divergence from the base model compared to other ablations. The model retains the Qwen3.5 architecture and its native 262,144-token context length.

Key Capabilities and Features

  • Censorship Ablation: Engineered to minimize content refusals while preserving the model's core reasoning abilities, with minimal deviation from the base model's original weights.
  • Distilled Chain-of-Thought: Incorporates a <think> block in its responses, directly learned from high-quality teacher traces, enhancing its reasoning process.
  • Mathematics and Code Emphasis: The distillation dataset was heavily weighted towards complex mathematical problems and competitive programming, making it proficient in these domains.
  • Native Function Calling: Supports function calling as per Qwen3.5 specifications without requiring additional wrappers or fine-tuning.
  • Full Fine-tune: All parameters were updated during the distillation process, not just adapters.

Performance Highlights

While the base Qwen3.8-9B showed a slight decrease in gsm8k_cot performance compared to Qwen3.5-9B, it demonstrated significant improvements in MMLU (CoT, 57 subjects), achieving an accuracy of 0.751 (flexible-extract) compared to 0.546 for the base, representing a +0.205 gain. This indicates strong general reasoning capabilities inherited from the teacher model.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

  • Complex Reasoning: Excels in tasks demanding detailed thought processes, especially in mathematics and coding.
  • Uncensored Content Generation: Designed for scenarios where reduced content refusal is critical, while maintaining a high degree of fidelity to the original model's output style.
  • Single GPU Deployment: Its 9B parameter size allows for efficient deployment on a single GPU, making it accessible for various applications.