insraq/Qwen3.5-2B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

insraq/Qwen3.5-2B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated is a 2.3 billion parameter causal language model, a decensored version of EmperoAI's Qwen3.8-2B. This model is a full-parameter distillation of Qwen3.8 2.4T A95B into the Qwen3.5-2B architecture, trained on curated chain-of-thought traces for reasoning and instruction following. It features a 32,768-token context length and native function calling, making it suitable for edge deployments requiring strong reasoning capabilities in a small footprint.

Loading preview...

Model Overview: insraq/Qwen3.5-2B-EmperoAI-Qwen3.8-Distill-Heretic-Abliterated

This model is a 2.3 billion parameter, decensored variant of EmperoAI's Qwen3.8-2B, built upon the Qwen3.5-2B architecture. Developed by Empero, it represents a full-parameter distillation of the larger Qwen3.8 2.4T A95B model, specifically trained on approximately 30,000 curated teacher traces focusing on dense chain-of-thought reasoning, mathematics, and instruction following. It inherits a substantial 262,144-token native context length from its Qwen3.5 base and supports native function calling.

Key Capabilities & Features

  • Distilled Chain-of-Thought: Learns directly from high-quality Qwen3.8 2.4T A95B traces, opening answers with a <think> block for enhanced reasoning.
  • Consistent Curriculum: Trained on the same quality-filtered trace mix as its larger 4B and 9B siblings, ensuring a consistent reasoning character.
  • Edge-Optimized: At 2.3B parameters, it's designed for efficient deployment on devices like phones, single-board computers, and CPU-only machines.
  • Decensored: This specific version has undergone "Abliteration" using Heretic v1.4.0, significantly reducing refusals (3/100 compared to 82/100 in the original).
  • Strong Reasoning Performance: Benchmarks show significant improvements over the Qwen3.5-2B base, with a +0.310 exact match increase on gsm8k_cot and a +0.265 accuracy increase on mmlu (CoT, flexible-extract).

Good for

  • Reasoning-intensive tasks: Excels in mathematical problem-solving and general reasoning due to its chain-of-thought distillation.
  • Instruction following: Designed to accurately follow complex instructions.
  • Edge deployments: Its small size (2B parameters) makes it suitable for resource-constrained environments.
  • Applications requiring native function calling: Supports Qwen3.5's function calling specification without additional fine-tuning.
  • Use cases requiring less refusal behavior: The decensored nature makes it more permissive compared to its base model.

Finetunes
1 models