coder3101/Qwen3.5-9B-heretic

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

coder3101/Qwen3.5-9B-heretic is a 9 billion parameter, decensored version of the Qwen/Qwen3.5-9B multimodal language model, created using the Heretic v1.2.0 tool for multi-directional refusal suppression. This model retains the original Qwen3.5's unified vision-language foundation, efficient hybrid architecture, and broad linguistic coverage, while significantly reducing refusal rates from 86/100 to 10/100. It is designed for applications requiring less restrictive content generation, excelling in multimodal understanding, reasoning, and agentic tasks with a native context length of 262,144 tokens.

Loading preview...

Overview

coder3101/Qwen3.5-9B-heretic is a 9 billion parameter multimodal language model, derived from Qwen/Qwen3.5-9B. Its primary distinction is being a decensored version, achieved through the application of the Heretic v1.2.0 tool with multi-directional refusal suppression. This modification drastically reduces the model's refusal rate from 86/100 to 10/100, as indicated by KL divergence and refusal metrics.

Key Capabilities

  • Decensored Content Generation: Significantly lower refusal rates compared to the base model, enabling broader content exploration.
  • Unified Vision-Language Foundation: Integrates multimodal tokens for strong performance across reasoning, coding, agentic tasks, and visual understanding.
  • Efficient Hybrid Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference with low latency.
  • Scalable RL Generalization: Enhanced real-world adaptability through reinforcement learning across diverse task distributions.
  • Global Linguistic Coverage: Supports 201 languages and dialects for inclusive, worldwide deployment.
  • Ultra-Long Context: Natively handles up to 262,144 tokens, extensible to 1,010,000 tokens using YaRN scaling techniques.
  • Tool Calling: Excels in agentic usage, compatible with frameworks like Qwen-Agent and Qwen Code.

Good for

  • Applications requiring less restrictive or unfiltered content generation.
  • Multimodal tasks involving complex reasoning, coding, and visual understanding.
  • Developing agents that interact with tools and environments.
  • Processing and generating content in a wide array of languages.
  • Handling extremely long documents or conversational histories.

Finetunes
1 models