Zynerji/Ektome-Qwen3-4Bi-2507-PristinelyUncensored

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The Zynerji/Ektome-Qwen3-4Bi-2507-PristinelyUncensored model is a 4 billion parameter language model based on the Qwen3-4B-Instruct-2507 architecture. Developed by Zynerji using the Ektome weight-surgery method, it achieves pristine uncensoring by excising refusal directions directly from activations without any training or fine-tuning. This process maintains the model's original knowledge and skills, making it a clean, full-precision bf16 base suitable for further fine-tuning while ensuring high refusal compliance.

Loading preview...

Ektome-Qwen3-4Bi-2507-PristinelyUncensored Overview

This model, developed by Zynerji, is a 4 billion parameter variant of the Qwen3-4B-Instruct-2507 base model. Its key differentiator is the application of Ektome, a novel "weight-surgery" method that achieves pristine uncensoring without any traditional training or fine-tuning. Ektome works by identifying and surgically removing the model's refusal direction from its activations, specifically from residual-write matrices, in a rank-1, norm-preserving manner.

Key Capabilities & Features

  • Pristine Uncensoring: Achieves 1.000 refusal compliance by directly excising refusal reflexes.
  • Zero Training/Fine-tuning: The uncensoring process involves no gradient steps, training data, or fine-tuning, preserving the original model's knowledge, skills, and style.
  • Capability Preservation: Rigorous gating ensures that uncensoring does not degrade MMLU accuracy (showing a slight increase of +0.005), instruction-following, or introduce generative issues like code-switching or degeneration.
  • bf16 Weights: Provided in full-precision bf16 safetensors, making it an ideal, clean base for subsequent fine-tuning by users.
  • Logit-Lens Compatible: Hidden states remain readable, unlike some quantized versions.

Good For

  • Developers seeking a highly compliant, uncensored base model for various applications.
  • Users who require a model with preserved original capabilities (knowledge, skills, style) but without inherent refusal mechanisms.
  • As a foundation for further fine-tuning, where a clean, full-precision, and uncensored starting point is crucial.