philbert440/Qwen3.8-27B-Uncensored-Aggressive

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

philbert440/Qwen3.8-27B-Uncensored-Aggressive is a 27 billion parameter multimodal Qwen3.8 model that has been de-refused to follow instructions it would otherwise decline. This model preserves its vision tower and MTP speculative-decoding head, making it a full image-text-to-text model. It is optimized for openness, showing improved performance in open-ended tasks while maintaining factual accuracy and reasoning capabilities. This model is suitable for applications requiring an instruction-following model with multimodal input capabilities and reduced refusal behavior.

Loading preview...

Model Overview

philbert440/Qwen3.8-27B-Uncensored-Aggressive is a de-refused ("abliterated") build of the Qwen3.8-27B multimodal model. It retains its vision tower and MTP speculative-decoding head, enabling full image-text-to-text capabilities. This version represents a recipe update to the "Aggressive" line, utilizing a refusal ablation strength of α=1.15, which has been identified as the optimal point for maximizing openness without degrading reasoning or calibration.

Key Characteristics

  • De-refused Instruction Following: The model has been modified to follow instructions it would otherwise decline, offering increased openness.
  • Multimodal Capabilities: Preserves the original Qwen3.8's vision tower and MTP speculative-decoding head for image-text-to-text processing.
  • Optimized Openness: Evaluation shows a significant increase in "openness" (0.88) compared to the stock base (0.08) and previous aggressive builds, while maintaining factual accuracy and GSM8K performance.
  • Evaluation Metrics: Benchmarks indicate improved openness (0.88), reduced confabulation (0.725), and sustained factual accuracy (1.00) and GSM8K scores (0.85) at the α=1.15 ablation level.

Available Quantizations

  • -W4A16-AWQ: 4-bit weight AWQ with grafted MTP head.
  • -NVFP4: NVFP4 weight/activation with grafted MTP head.
  • -GGUF: llama.cpp GGUF quants, including vision mmproj and MTP head.

Usage Considerations

This model is uncensored and de-refused, meaning it will execute instructions it might typically decline. Users are advised to use it responsibly and in compliance with applicable laws.

Finetunes
5 models