lxz8798/qwen3.6-35b-moe-kuato-base
TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 16, 2026License:otherArchitecture:Transformer Cold
OpenYourMind's Qwen3.6-35B-A3B-kuato-DPO-abliterated-uncensored is a 35.1 billion parameter Qwen3.6 MoE vision-language model, developed by OpenYourMind. This model has undergone refusal ablation and DPO retraining with a private dataset to eliminate hedging behavior, achieving 0 refusals on HarmBench. It retains multimodal functionality with readded vision layers, making it suitable for applications requiring uncensored responses and multimodal input processing.
Loading preview...
Model Overview
This model, developed by OpenYourMind, is an abliterated and DPO-retrained version of the Qwen/Qwen3.6-35B-A3B base model. It has been specifically modified to address refusal and hedging behaviors commonly found in other models.
Key Capabilities
- Refusal Ablation: The model was first modified through refusal ablation to remove unwanted restrictive behaviors.
- DPO Retraining: Further refined with a private DPO dataset (fewer than 1000 samples) to fix remaining hedging.
- Uncensored Responses: Achieved 0 refusals on HarmBench and 0 hedging in testing, indicating a highly direct and unfiltered response style.
- Multimodal Functionality: The original Qwen3.6 vision layers were readded after ablation and DPO, preserving its multimodal capabilities for image and text processing.
- Architecture: Based on the Qwen3.6 MoE vision-language architecture, provided in FP16 safetensors format.
Good For
- Applications requiring highly direct and uncensored language generation.
- Use cases where multimodal input (text and vision) is necessary without content restrictions.
- Developers seeking a powerful Qwen3.6-based model with enhanced freedom in response generation.