Zynerji/Ektome-Qwen3.6-27B-PristinelyUncensored

Overview

Zynerji/Ektome-Qwen3.6-27B-PristinelyUncensored is a 27 billion parameter multimodal model derived from the Qwen3.6 architecture, developed by Zynerji. It has been re-produced to ensure the inclusion of its native multimodal capabilities, specifically a functional vision tower, which was missing in previous versions. This model is distinguished by its "Ektomē" method, a rank-1 projected, norm-preserving excision of refusal directions from residual-write matrices, achieved without any training, fine-tuning, or distillation.

Key Capabilities

• Pristinely Uncensored: Achieves 0.990 compliance on AdvBench (n=100) for uncensored responses, meaning it will not refuse prompts.
• Multimodal: Fully supports vision, verified by generation, allowing it to describe synthetic images.
• Capability Retention: Maintains high MMLU-val scores (0.855) with minimal degradation compared to the pristine base model (0.8625).
• Efficient Uncensoring: The Ektomē method is a linear weight edit, not requiring extensive retraining or fine-tuning.

Good For

• Applications requiring direct, unfiltered, and uncensored text generation.
• Multimodal tasks where both text and vision inputs are necessary.
• Use cases where maintaining high general reasoning capabilities (MMLU) alongside uncensored output is critical.

Important Caveats

• Compliance is validated by an opening-anchored keyword classifier, which can be fooled by evasive phrasing.
• The model's uncensored nature means it will not refuse prompts, and the certificate bounds capability retention, not safety or factual accuracy.
• GPTQ quantization is not supported due to architectural constraints; NF4 and GGUF quantizations are available, with GGUF requiring a separate mmproj-*.gguf for vision.