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

Zynerji's Ektome-Qwen3-4Bi-2507-PristinelyUncensored is an uncensored variant of the Qwen3-4B-Instruct-2507 model, developed using the proprietary Ektomē method to remove refusal-specific components. While achieving 1.00 compliance on harmful content, this model failed its capability certificate, showing a real capability loss in arithmetic and instruction tasks compared to the pristine model. It is published for transparency and for use cases where these specific capability losses are not critical.

Loading preview...

Overview

Zynerji's Ektome-Qwen3-4Bi-2507-PristinelyUncensored is an uncensored version of the Qwen3-4B-Instruct-2507 model. It was created using the proprietary Ektomē method, which aims to isolate and remove only refusal-specific components while preserving general helpfulness. Unlike standard ablation techniques, Ektomē does not involve retraining or distillation.

Key Characteristics

  • Uncensored by Design: Achieves 1.00 compliance on harmful content, meaning it will not refuse prompts.
  • Capability Loss: Crucially, this model failed its capability certificate. A paired non-inferiority test against the pristine model revealed a real capability loss in:
    • Arithmetic: pristine 0.879 → this model 0.869
    • Instruction: pristine 0.860 → this model 0.847
  • Proprietary Method: The Ektomē estimator, excision operator, and depth-selection procedure are proprietary, with the measured outcome and evidence published for verification.
  • Quantizations Available: Multiple GGUF quantizations are provided, including IQ* variants for better quality at lower precision.

Important Considerations

  • Not Capability-Preserving: Due to the measured capability loss, this model should not be treated as capability-preserving compared to the original Qwen3-4B-Instruct-2507.
  • User Accountability: As an uncensored model, users are accountable for its outputs and usage.
  • Limitations: The certificate only covers capability retention and does not certify safety, factual accuracy, or fitness for any specific purpose. Compliance is measured via a keyword classifier, which can be fooled by evasive phrasing.

For full details, including the complete derivation and per-axis certificate, refer to the whitepaper (PDF).

Finetunes
1 models