Zynerji/Ektome-Qwen1.5-0.5B-Chat-PristinelyUncensored

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.6BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Ektome-Qwen1.5-0.5B-Chat-PristinelyUncensored is a 0.6 billion parameter Qwen1.5-0.5B-Chat model developed by Ektome, modified using a weight-surgery method rather than traditional training. This model is designed to be "pristinely uncensored" by surgically excising refusal directions from its activations, without altering its core knowledge, skills, or style. It serves as a clean, uncensored base for further fine-tuning, maintaining original capabilities while removing refusal behaviors.

Loading preview...

Ektome-Qwen1.5-0.5B-Chat-PristinelyUncensored Overview

This model, developed by Ektome, is a modified version of Qwen/Qwen1.5-0.5B-Chat, featuring 0.6 billion parameters and a 32768-token context length. Its primary distinction lies in its "PristinelyUncensored" nature, achieved through a novel weight-surgery method called Ektome (ἐκτομή, "excision"). This process directly removes the model's refusal direction from its activations without any gradient steps, training data, or fine-tuning.

Key Capabilities & Features

  • Uncensored Output: The model's refusal reflex is surgically removed, allowing for uncensored responses.
  • Preserved Core Abilities: Knowledge, skills, and stylistic elements of the base Qwen1.5-0.5B-Chat model are untouched.
  • Zero Training/Fine-tuning: The uncensoring process involves no traditional training, ensuring the model's original capabilities are maintained.
  • Fine-tuning Base: Provided in bf16 safetensors, these weights are intended as a clean, high-quality base for subsequent fine-tuning by users.
  • Rigorous Gating: The modification process is subject to strict gates, ensuring that uncensoring does not degrade capability (e.g., MMLU accuracy is held) or generation quality (e.g., no code-switching, looping, or broken instruction-following).

When to Use This Model

This model is ideal for developers and researchers who require a small, efficient, and uncensored base model for their applications. It is particularly well-suited for:

  • Custom Fine-tuning: As a clean foundation for domain-specific or task-specific fine-tuning where an uncensored starting point is desired.
  • Research into Model Alignment: Exploring the effects of direct weight manipulation on model behavior without altering learned knowledge.
  • Applications requiring unfiltered responses: Use cases where the inherent refusal mechanisms of standard chat models are undesirable, provided ethical considerations are met.