ajgazin/Swift-Qwen3.8-27B-Uncensored-MTP

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 15, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

ajgazin/Swift-Qwen3.8-27B-Uncensored-MTP is a 27 billion parameter language model derived from UkisAI's Swift-Qwen3.8-27B, which itself is a reasoning-efficient fine-tune of Qwen3.8-27B. This model applies a single-direction refusal ablation technique, similar to orcarouter/Qwen3.8-27B-Uncensored, to significantly reduce refusals while preserving the vision tower and the MTP (Multi-Turn Prediction) head for self-speculative decoding. It is designed to answer requests that the original Swift model might decline, making it suitable for use cases requiring less restrictive content generation.

Loading preview...

Overview

This model, ajgazin/Swift-Qwen3.8-27B-Uncensored-MTP, is a 27 billion parameter language model built upon UkisAI's Swift-Qwen3.8-27B. Its primary differentiator is the application of a refusal ablation technique, inspired by orcarouter/Qwen3.8-27B-Uncensored, which significantly reduces the model's tendency to refuse prompts. This "uncensoring" process involves projecting a specific refusal direction out of Swift's matrices, allowing it to generate responses to queries that the original Swift model would typically decline.

Key Capabilities

  • Reduced Refusals: Achieves a refusal rate of 15/100 prompts on mlabonne/harmful_behaviors compared to 98/100 for the base Swift model, as measured by Heretic's keyword-based detector.
  • Preserved Vision Tower: The model retains the original vision tower from its base, allowing for multimodal image-to-text capabilities.
  • MTP Head for Speculative Decoding: The Multi-Turn Prediction (MTP) head is kept and consistently edited, enabling efficient self-speculative decoding for faster inference.
  • Compatibility: Maintains the architecture, tokenizer, and chat template of Swift and Qwen3.8-27B, ensuring broad compatibility with existing tools and workflows.

Good for

  • Unrestricted Content Generation: Ideal for applications where the base model's refusal behavior is undesirable, and a more open-ended response generation is required.
  • Research into Refusal Mechanisms: Useful for studying the effects of refusal ablation techniques on large language models.
  • Multimodal Applications: Can be used in scenarios requiring both text and image processing, leveraging its preserved vision capabilities.
  • Efficient Inference: Benefits from the MTP head for self-speculative decoding, potentially offering faster response times in deployment.

Finetunes
1 models