KridgeDookie/Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

KridgeDookie/Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS is a 27 billion parameter multimodal derivative of Qwen/Qwen3.8-27B, developed by KridgeDookie. This model is specifically engineered to sharply reduce refusal behavior, achieving 0 refusals across extensive internal prompt screens. It retains the upstream hybrid text-and-vision backbone, making it suitable for applications requiring robust, uncensored responses in both text and multimodal contexts.

Loading preview...

Model Overview

KridgeDookie/Qwen3.8-27B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS is a 27 billion parameter multimodal model derived from Qwen/Qwen3.8-27B. Its primary modification focuses on significantly reducing refusal behavior, making it highly "uncensored" in its responses. The model maintains the original Qwen3.8's hybrid text-and-vision architecture, supporting both text-only and multimodal inputs.

Key Capabilities & Performance

  • Refusal Reduction: Achieved 0 refusals across an internal 842-prompt screen and a separate 126-prompt family holdout, demonstrating high usability.
  • Coherence: Passed 23 out of 24 coherence checks across tasks including coding, JSON, debugging, explanation, and mathematics.
  • Multimodal Support: Retains the ability to process and respond to vision inputs, validated by a successful multimodal reload smoke test.
  • BF16 Precision: The primary Transformers checkpoint is in BF16, offering a balance of performance and memory usage.

Use Cases & Considerations

This model is particularly well-suited for applications where minimizing refusal behavior is critical, such as creative writing, open-ended dialogue, or scenarios requiring direct answers without filtering. While it excels in reducing refusals, users should note that this does not guarantee factual accuracy or preservation of upstream reasoning, factuality, coding, or vision benchmark scores. The model is available in safetensors for multimodal use and GGUF formats for local text-generation, with Qwen3.8's thinking mode still accessible.