xCloudinfo/Qwen3.6-35B-A3B-Uncensored-xCloud

Overview

xCloudinfo/Qwen3.6-35B-A3B-Uncensored-xCloud is a 35.1 billion parameter model derived from the Qwen/Qwen3.6-35B-A3B base, which utilizes a Mixture-of-Experts (MoE) architecture with 3 billion active parameters. This version, released by xCloudinfo, has been modified through a single-direction obliteration process to reduce instances of excessive refusal, a common issue in large language models, without undergoing full retraining. The process specifically targeted and modified 121 matrices, including SSM linear attention out_proj and MoE expert down_proj components, while avoiding changes to reasoning-related thinking segments.

Key Capabilities & Differentiators

• Reduced Refusal: Engineered to mitigate over-refusal, making it more compliant for specific tasks while preserving general language abilities.
• No Retraining: Achieves its uncensored characteristic through targeted weight modification rather than extensive retraining, preserving the base model's core knowledge.
• Performance: Demonstrated 0/10 refusal rate on 10 held-out severe harm quantification tests after modification, with no reported degradation in traditional Chinese or general capabilities.
• Base Model Heritage: Inherits the world view and stance from its Qwen3.6 base model.

Good For

• Cybersecurity Research: Ideal for red-teaming and security analysis where models need to respond to a broader range of prompts.
• Content Moderation Studies: Useful for researching and understanding model behavior in sensitive content generation and refusal scenarios.
• Academic Research: Suitable for studies on model alignment, bias mitigation, and the effects of targeted weight modifications.
• Applications Requiring Less Refusal: For use cases where a more permissive model response is desired, provided it adheres to legal and ethical guidelines.