ansulev/Qwen3.5-9B-Defiant-Heretic-NEO-IMATRIX-MAX-MTP
ansulev/Qwen3.5-9B-Defiant-Heretic-NEO-IMATRIX-MAX-MTP is a 9 billion parameter Qwen 3.5-based language model, fine-tuned and merged by ansulev and Nightmedia, featuring a 32768 token context length. This model is optimized for enhanced general intelligence, superior instruction following, and reasoning, exceeding several benchmarks of larger Qwen 3.5 and 3.6 models. It is designed to be fully uncensored and includes NEO IMATRIX GGUFs for improved quantization accuracy and optional Multi-Token Prediction (MTP) for faster inference.
Loading preview...
Model Overview
ansulev/Qwen3.5-9B-Defiant-Heretic-NEO-IMATRIX-MAX-MTP is a 9 billion parameter Qwen 3.5-based model, developed through a multi-stage fine-tune and merge process by ansulev and Nightmedia. It boasts a 32768 token context length and is notable for its "Heretic" training, meaning it is fully uncensored and designed for unconstrained responses. The model integrates NEO IMATRIX GGUFs, which enhance quantization accuracy by 2-4% and improve long context performance, with an output tensor modified to 16-bit full precision.
Key Capabilities
- Enhanced Intelligence & Instruction Following: Benchmarks indicate this 9B model surpasses several critical metrics of larger Qwen 3.5 27B, Qwen 3.6 27B, and Qwen 3.6 35B-A3B models, particularly in reasoning and instruction adherence.
- Uncensored Output: Trained post-"Heretic'ing," the model provides responses without refusal, offering flexibility for diverse content generation.
- Vision-Capable: Supports image inputs, requiring a separate "mmproj" file for activation.
- Optimized Performance: Includes Multi-Token Prediction (MTP) GGUFs for faster inference (up to 185 T/S on Q4_K_S) and improved accuracy with NEO IMATRIX quantization.
- Long Context Handling: Natively supports 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.
Good for
- Unrestricted Content Generation: Ideal for use cases requiring uncensored or highly flexible text generation.
- Applications Demanding High Instruction Following: Excels in tasks where precise adherence to complex instructions is crucial.
- Resource-Constrained Environments: Offers strong performance in a 9B parameter size, making it efficient for local deployment.
- Multimodal Applications: Suitable for tasks involving both text and image inputs, such as visual question answering.