SassyDiffusion/Qwen3.5-9B-heretic
SassyDiffusion/Qwen3.5-9B-heretic is a 9 billion parameter causal language model, a decensored version of Qwen/Qwen3.5-9B created using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method. It features a native context length of 262,144 tokens, extensible up to 1,010,000 tokens via YaRN scaling, and excels in multimodal learning, supporting unified vision-language inputs including images and video. This model is optimized for general agent tasks, reasoning, coding, and multilingual understanding across 201 languages, demonstrating significantly reduced refusals compared to its original counterpart.
Loading preview...
Overview
SassyDiffusion/Qwen3.5-9B-heretic is a 9 billion parameter multimodal causal language model, derived from Qwen/Qwen3.5-9B. This version has been decensored using the Heretic v1.2.0 tool with the Arbitrary-Rank Ablation (ARA) method, specifically targeting layers 7 through 18. A key differentiator is its significantly reduced refusal rate (7/100) compared to the original model (93/100), while maintaining a low KL divergence of 0.0168.
Key Capabilities
- Unified Vision-Language Foundation: Achieves strong performance across reasoning, coding, agent tasks, and visual understanding by integrating multimodal tokens early in training.
- Efficient Hybrid Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference with minimal latency.
- Scalable RL Generalization: Enhanced real-world adaptability through reinforcement learning scaled across million-agent environments.
- Global Linguistic Coverage: Supports 201 languages and dialects, offering nuanced cultural and regional understanding.
- Extended Context Length: Natively handles up to 262,144 tokens, extensible to 1,010,000 tokens using YaRN scaling techniques.
- Multimodal Input: Capable of processing text, image, and video inputs, making it suitable for complex multimodal tasks.
What Makes This Model Different?
This model's primary distinction lies in its decensored nature, achieved through the Heretic tool. This modification aims to provide responses with significantly fewer refusals, as evidenced by the reported metrics. While retaining the robust multimodal and multilingual capabilities of the base Qwen3.5-9B, the 'heretic' version offers a more unconstrained output behavior, which can be critical for specific applications where content filtering is undesirable.
Should I Use This for My Use Case?
- Good for:
- Applications requiring unfiltered or less constrained responses where the original Qwen3.5-9B might refuse.
- Multimodal tasks involving image and video understanding alongside text.
- Multilingual applications needing broad language support.
- Agentic workflows and tool calling, especially with Qwen-Agent or Qwen Code.
- Tasks requiring long context processing (up to 1M tokens).
- Consider alternatives if:
- Strict content moderation and safety filters are paramount.
- Your application does not require multimodal capabilities or extensive multilingual support, and a smaller, more specialized model might be more efficient.