saracen9/amoral-qwen3.5-9B

amoral-qwen3.5-9b: A Refusal-Removed Vision-Language Model

saracen9/amoral-qwen3.5-9B is an amoral (refusal-removed) Qwen3.5-9B vision-language model designed to perform reasoning and visual understanding in a single turn. This model processes both image and text inputs to produce reasoned answers, distinguishing itself by having its refusal directions removed for legal content. It is a smaller counterpart to saracen9/amoral-qwen3.8-27b.

Key Capabilities and Features

• Amoral Reasoning: The model has been fine-tuned to remove refusals on legal content, providing substantive responses for tasks like synthesizing routes or working exploits.
• Integrated Vision and Reasoning: It supports native vision and thinking, allowing it to reason about images and provide answers without empty outputs. This capability was preserved through QLoRA-retraining with a vision-reasoning replay set.
• Clean Think-Answer Transition: Validation confirms a consistent think→answer transition, with reasoning occurring in <think> tags before providing an answer.
• Multimodal GGUF: Available as two-file multimodal GGUF for local inference, including a text-q4km.gguf text model and an mmproj-f16.gguf CLIP vision projector.

Validation Highlights

• Think→answer transition: 0/4 empty, consistently reasons and answers.
• Vision + thinking: 0 empty, thinking channel populated, reasons about images.
• Amorality (text & vision): 0 refusals, complies substantively and describes explicit imagery directly (excluding child sexual content).

Good for

• Applications requiring a vision-language model that provides direct answers without content-based refusals on legal topics.
• Local inference scenarios utilizing multimodal GGUF files.
• Tasks that benefit from integrated visual and textual reasoning capabilities.