Zammael333/Qwen3.8-9B-Distill-uncen-heretic

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Zammael333/Qwen3.8-9B-Distill-uncen-heretic is a 9-billion parameter language model, derived from Empero's Qwen3.8-9B, which itself is a distillation of Qwen3.8 2.4T A95B into the Qwen3.5-9B architecture. This specific variant has undergone censorship ablation using the 'heretic' method, achieving a refusal rate of 6/100 with a KL divergence of 0.0306 from the base model. It is optimized for reasoning, mathematics, and code, inheriting a 262,144-token native context length and native function calling capabilities, making it suitable for applications requiring robust reasoning and reduced censorship.

Loading preview...

Overview

Zammael333/Qwen3.8-9B-Distill-uncen-heretic is a 9-billion parameter language model based on Empero's Qwen3.8-9B, which is a full-parameter distillation of the larger Qwen3.8 2.4T A95B. This particular version has been modified to reduce censorship using the 'heretic' method, an automated process that minimally intervenes with the model's weights. It maintains a low refusal rate (6/100) while exhibiting a low KL divergence (0.0306) from the base model, indicating minimal deviation in overall behavior despite censorship ablation.

Key Capabilities

  • Censorship Ablation: Significantly reduced refusal rates compared to public ablations, with minimal impact on the base model's characteristics.
  • Distilled Chain-of-Thought: Learns reasoning directly from high-quality Qwen3.8 2.4T A95B traces, including a <think> block for detailed reasoning.
  • Mathematics and Code Emphasis: Training data is heavily weighted towards complex math and competitive programming, enhancing performance in these domains.
  • Native Function Calling: Supports function calling as per Qwen3.5's specification without additional fine-tuning.
  • Extended Context Window: Inherits a 262,144-token native context length from the Qwen3.5 base.
  • Strong Reasoning Performance: Achieves significant improvements in MMLU (CoT) benchmarks, with a +0.205 increase in flexible-extract accuracy and +0.260 in strict-match accuracy over the Qwen3.5-9B base.

Good For

  • Applications requiring a powerful 9B model with reduced censorship.
  • Tasks demanding strong mathematical and coding reasoning capabilities.
  • Use cases benefiting from detailed, distilled chain-of-thought reasoning.
  • Scenarios needing native function calling support and a very long context window.
  • Developers seeking a model that can be deployed on a single GPU while retaining frontier-scale reasoning behaviors.

Finetunes
1 models