d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-BF16
d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-BF16 is a 27 billion parameter, BF16 uncensored derivative of ukisai/Swift-1.5-Qwen3.8-27b, developed by d0xin. It utilizes rank-1 directional residual-stream ablation to reduce refusal behavior while preserving the original Swift 1.5's reasoning-efficient, coding, agentic, tool-calling, multimodal, and long-context capabilities. With a configured context length of 262,144 tokens, this model is designed for applications requiring robust performance across complex tasks with reduced content restrictions.
Loading preview...
Overview
d0xin/Swift-1.5-Qwen3.8-27B-Uncensored-BF16 is a 27 billion parameter model derived from ukisai/Swift-1.5-Qwen3.8-27b. This model is an independent, uncensored BF16 checkpoint, specifically modified to reduce refusal behavior while retaining the advanced capabilities of its Swift 1.5 base. It leverages rank-1 directional residual-stream ablation for this transformation.
Key Capabilities
- Reduced Refusal Behavior: Intentionally modified to generate content that the upstream model might refuse, offering greater flexibility.
- Preserved Core Strengths: Maintains Swift 1.5's strong performance in reasoning, coding, agentic workflows, tool-calling, and multimodal tasks.
- Long Context Support: Configured for an extensive context length of 262,144 tokens, suitable for complex, long-horizon tasks.
- Multimodal Architecture: Retains the Qwen3.8 multimodal architecture and vision tower, which remains unchanged by the ablation process.
- Tool-Calling: Supports tool-calling functionalities, leveraging the
qwen3_coderparser.
Good For
- Applications requiring an uncensored or refusal-reduced language model.
- Workloads demanding high reasoning efficiency and coding proficiency.
- Agentic applications and tasks involving long-horizon planning.
- Use cases benefiting from multimodal input processing.
- Developers seeking a model with extensive context handling for complex prompts.