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

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 Cold

d0xin/Swift-Qwen3.8-27B-Uncensored-BF16 is a 27 billion parameter, BF16 checkpoint model derived from UkisAI's Swift-Qwen3.8-27B, featuring a 262,144 token context length. This independent derivative has been modified using rank-1 directional residual-stream ablation to significantly reduce refusal behavior, achieving 0/100 refusals on a fixed evaluation set. It preserves the original Swift-Qwen3.8's reasoning, agentic, tool-calling, multimodal, and long-context capabilities while offering an uncensored output profile.

Loading preview...

Overview

d0xin/Swift-Qwen3.8-27B-Uncensored-BF16 is a 27 billion parameter model, a BF16 checkpoint derived from UkisAI's Swift-Qwen3.8-27B. Its primary distinction is the removal of refusal behavior through rank-1 directional residual-stream ablation, achieving a 0/100 refusal rate on a fixed 100-prompt evaluation. This modification aims to retain the original model's advanced capabilities while providing an uncensored output.

Key Capabilities

  • Uncensored Output: Intentionally modified to reduce refusal behavior, allowing for broader content generation.
  • High Context Length: Supports a configured maximum position length of 262,144 tokens.
  • Preserved Intelligence: Validation against the original Swift BF16 showed no measurable intelligence degradation, with a slight improvement of +1.68 percentage points on a 298-example comparison.
  • Multimodal Support: Retains the Qwen multimodal architecture and vision tower, with vision tensors unchanged by the ablation process.
  • Tool Calling: Supports OpenAI-style tools requests and integrates with SGLang's qwen3_coder tool-call parser.
  • Reasoning & Agentic Workloads: Inherits Qwen3.8's reasoning capabilities, with configurable reasoning_effort levels (xhigh, medium, low) for complex tasks.

Good For

  • Applications requiring an uncensored language model that maintains strong reasoning, agentic, and multimodal capabilities.
  • Developers needing a model with a very long context window for complex tasks.
  • Research and experimentation where reduced refusal behavior is desired, provided users ensure responsible and lawful use.

Finetunes
1 models