llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved

TEXT GENERATIONConcurrency Cost:3Model Size:35.1BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved is a 35.1 billion parameter causal language model based on the Qwen3.6-35B-A3B architecture, developed by Qwen and further decensored by llmfan46. This model significantly reduces refusals by 88% (10/100 vs 83/100) while maintaining original model quality with a 0.0015 KL divergence. It is optimized for agentic coding, preserving reasoning context, and supports a native context length of 32,768 tokens, extensible up to 1,010,000 tokens with YaRN scaling.

Loading preview...

Model Overview

This model, llmfan46/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved, is a 35.1 billion parameter variant of the Qwen3.6-35B-A3B causal language model. It has been decensored using the Heretic v1.3.0 method with Magnitude-Preserving Orthogonal Ablation (MPOA), resulting in an 88% reduction in refusals (10/100 compared to 83/100 for the original) while preserving model quality with a low KL divergence of 0.0015.

Key Capabilities

  • Reduced Refusals: Significantly lowers content restrictions without compromising the model's baseline performance.
  • Agentic Coding: Enhanced capabilities for handling frontend workflows and repository-level reasoning.
  • Thinking Preservation: Retains reasoning context from historical messages, improving iterative development and reducing overhead.
  • Multimodal Support: Processes text, image, and video inputs, making it suitable for diverse applications.
  • Extended Context Length: Natively supports 32,768 tokens, extensible up to 1,010,000 tokens using YaRN scaling for ultra-long texts.
  • Tool Use: Excels in tool calling, with recommended integration via Qwen-Agent and Qwen Code for terminal-based AI agent applications.

Good For

  • Developers requiring uncensored outputs: Ideal for applications where content restrictions are undesirable.
  • Agentic workflows: Particularly strong in coding agents, general agents, and tool-use scenarios.
  • Complex problem-solving: Benefits from thinking preservation and extended context for detailed reasoning in tasks like math and programming competitions.
  • Multimodal applications: Suitable for tasks involving image and video understanding alongside text.

Finetunes
5 models