Avesed/Qwopus3.6-27B-v2-abliterated
Avesed/Qwopus3.6-27B-v2-abliterated is a 27 billion parameter Qwen3.5-hybrid (GatedDeltaNet linear-attention + gated full-attention) vision-language reasoning model. Developed by Avesed, this model has undergone refusal-ablation, reducing its refusal rate from 100% to 8% while preserving general capabilities. It excels in coding and mathematical reasoning, achieving 95.1% on HumanEval pass@1 and 86.0% on GSM8K. This model is suitable for applications requiring robust reasoning without excessive refusal behaviors.
Loading preview...
Qwopus3.6-27B-v2-abliterated Overview
This model is a 27 billion parameter variant of the Qwen3.5-hybrid architecture, featuring a GatedDeltaNet linear-attention and gated full-attention mechanism, originally from Jackrong/Qwopus3.6-27B-v2. Its primary distinction is the "abliteration" process, where refusal-direction orthogonalization was applied at layer 26 to the residual-stream write matrices (o_proj / down_proj). This method significantly reduced the model's refusal rate from 100% to 8% without fine-tuning, ensuring general capabilities are preserved.
Key Capabilities & Performance
- Refusal Reduction: Achieves a substantial drop in refusal rate (100% to 8%) through a novel orthogonalization method.
- Strong Reasoning: Demonstrates high performance across various benchmarks, including:
- HumanEval pass@1: 95.1%
- GSM8K: 86.0%
- MMLU-Pro: 83.2%
- Vision-Language: Based on a Qwen3.5 hybrid, indicating strong vision-language reasoning capabilities, with the vision tower untouched during ablation.
- Multi-Token-Prediction (MTP) Head: Includes an abliterated MTP head for speculative decoding, enhancing efficiency.
Use Cases
This model is particularly well-suited for applications requiring a powerful vision-language model with significantly reduced refusal behaviors, making it ideal for tasks demanding reliable and direct responses in coding, mathematical problem-solving, and general reasoning.