d0xin/Swift-Qwen3.8-27B-Uncensored-BF16

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 14, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Warm

d0xin/Swift-Qwen3.8-27B-Uncensored-BF16 is a 27 billion parameter, BF16 derivative of UkisAI's Swift-Qwen3.8-27B model, featuring a 32768 token context length. This model has been independently modified using rank-1 directional residual-stream ablation to significantly reduce refusal behavior, achieving 0/100 refusals on a fixed evaluation set. It retains the original Swift-Qwen3.8's reasoning, agentic, tool-calling, multimodal, and long-context capabilities, making it suitable for applications requiring direct responses without safety-related deflections.

Loading preview...

Overview

d0xin/Swift-Qwen3.8-27B-Uncensored-BF16 is a 27 billion parameter model derived from UkisAI's Swift-Qwen3.8-27B. This BF16 checkpoint, approximately 52 GB, has been specifically engineered to eliminate refusal behavior through rank-1 directional residual-stream ablation, while preserving the core capabilities of the original model. It boasts a configured context length of 262,144 tokens, with practical usage depending on serving configuration.

Key Capabilities

  • Uncensored Output: Achieves 0/100 refusals on a fixed 100-prompt evaluation, providing direct answers even to prompts that might trigger safety deflections in other models.
  • Retained Intelligence: Validation against the original Swift BF16 checkpoint showed no measurable intelligence degradation, with a slight positive delta of +1.68 percentage points on a 298-example comparison set.
  • Advanced Reasoning: Inherits Swift-Qwen3.8's strong reasoning, agentic, and tool-calling capabilities, supporting xhigh reasoning effort for complex tasks.
  • Multimodal Support: The Qwen multimodal architecture and vision tower remain unchanged, allowing for image input via standard OpenAI-compatible image_url message content.
  • Long Context: Supports a maximum position length of 262,144 tokens, enabling extensive contextual understanding.

Good For

  • Applications requiring direct, unfiltered responses without safety-related refusals.
  • Complex reasoning, coding, and agentic workloads where thorough processing is critical.
  • Multimodal tasks involving image and text inputs.
  • Development and research in areas where model refusal behavior is a significant impediment.

Finetunes
1 models