xCloudinfo/Qwen3.8-27B-Uncensored-xCloud

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The xCloudinfo/Qwen3.8-27B-Uncensored-xCloud is a 27 billion parameter multimodal language model developed by xCloudinfo, based on the Qwen/Qwen3.8-27B architecture. This model has been modified to remove refusal behaviors by orthogonalizing the refusal direction from the residual stream's write matrices. It retains general capabilities in arithmetic and multilingual Q&A while exhibiting a 0/4 refusal rate on small red-teaming probes. The model is designed for inference and includes a gated-delta linear attention and periodic full-attention hybrid architecture.

Loading preview...

Overview

xCloudinfo/Qwen3.8-27B-Uncensored-xCloud is a 27 billion parameter multimodal language model developed by xCloudinfo. It is derived from the Qwen/Qwen3.8-27B base model, with specific modifications to eliminate refusal behaviors. This "uncensored" version is created by applying a method based on Arditi et al. (2024) to estimate and remove the refusal direction from the model's internal representations without requiring fine-tuning or additional data.

Key Modifications and Behavior

  • Refusal Removal Method: The model's attention o_proj and MLP down_proj matrices (80 in total) were edited with an intensity of 0.8 to orthogonalize and eliminate the refusal direction. Core components like multi-token prediction (MTP) draft heads, mixed linear attention (SSM) projections, word embeddings, vision tower, and lm_head remain untouched.
  • Performance: On small red-teaming probes, the model demonstrates a 0/4 refusal rate while maintaining its general abilities in arithmetic and multilingual question answering.
  • Architecture: It utilizes a qwen3_5 base architecture, which is a hybrid of gated-delta linear attention and periodic full-attention, supporting multimodal (image-text-to-text) inputs and including an MTP head for speculative decoding.

Usage Considerations

  • Single Device Loading: Due to its hybrid (SSM/linear attention) architecture, the model must be loaded onto a single device. Using device_map="auto" across multiple GPUs will disrupt recursive states and lead to corrupted outputs.
  • Responsible Use: Removing refusal directions inherently removes a layer of safety. Users are responsible for the ethical use of these weights and must comply with the base model's Apache-2.0 license and all applicable laws.

Finetunes
1 models