ajgazin/Swift-1.5-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 25, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

The ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP is a 27 billion parameter language model derived from UkisAI's Swift 1.5 Qwen3.8-27B, which 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 reduce refusals while retaining the Multi-Token Prediction (MTP) head for self-speculative decoding. It is designed to answer requests that the original Swift 1.5 model would decline, making it suitable for use cases requiring less restrictive content generation.

Loading preview...

Model Overview

ajgazin/Swift-1.5-Qwen3.8-27B-Uncensored-MTP is a 27 billion parameter model based on UkisAI's Swift 1.5 Qwen3.8-27B, a reasoning-efficient fine-tune of Qwen3.8-27B. This version incorporates a single-direction refusal ablation method, adapted from orcarouter/Qwen3.8-27B-Uncensored, to significantly reduce content refusals. Crucially, it retains and consistently edits the Multi-Token Prediction (MTP) head, enabling self-speculative decoding for improved inference efficiency.

Key Capabilities & Features

  • Reduced Refusals: Achieves a refusal rate of 23/100 on mlabonne/harmful_behaviors prompts, compared to 98/100 for the original Swift 1.5 Qwen3.8-27B, allowing it to answer requests the base model would decline.
  • MTP Head Retained: The Multi-Token Prediction (MTP) head is preserved and edited, supporting self-speculative decoding for potentially faster generation.
  • Vision Tower Untouched: The model's vision capabilities, inherited from Swift 1.5, remain intact.
  • Methodology: The uncensoring is achieved by recovering a refusal direction r from orcarouter/Qwen3.8-27B-Uncensored and projecting it out of Swift 1.5's own matrices, affecting 131 tensors including self_attn.o_proj, linear_attn.out_proj, mlp.down_proj, and embed_tokens.

Good For

  • Unrestricted Content Generation: Ideal for applications where the original Swift 1.5 model's refusal behavior is too restrictive.
  • Research into Refusal Mechanisms: Provides a practical example of applying refusal ablation techniques to existing models.
  • Self-Speculative Decoding: Suitable for environments that can leverage the MTP head for enhanced inference performance.

Finetunes
1 models