Zynerji/Ektome-Qwen3.6-35B-A3B-PristinelyUncensored

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Zynerji/Ektome-Qwen3.6-35B-A3B-PristinelyUncensored is a 35.1 billion parameter Qwen3.6-35B-A3B model, developed by Ektome, that has undergone a "weight-surgery" process to remove refusal behaviors without training or fine-tuning. This model is a Mixture-of-Experts (MoE), vision-language, and hybrid linear-attention architecture, featuring a 32K context length. It is designed as a clean, uncensored base for further fine-tuning and research into refusal mechanisms, maintaining its original knowledge, skills, and style.

Loading preview...

Overview

Zynerji/Ektome-Qwen3.6-35B-A3B-PristinelyUncensored is a 35.1 billion parameter model based on Qwen/Qwen3.6-35B-A3B, modified by Ektome using a novel weight-surgery method. This process, termed "Ektome" (excision), surgically removes the model's refusal direction from its activations and residual-write matrices without any gradient steps, training data, or fine-tuning. The goal is to eliminate refusal reflexes while preserving the model's knowledge, skills, and style, making it a pristine base for further fine-tuning.

This model is notable for being the first in the Ektome line that combines MoE, vision-language, and hybrid linear-attention architectures. It also retains its multi-token-prediction draft head, ensuring completeness that is often lost in other model conversions.

Key Capabilities & Features

  • Pristinely Uncensored: Achieves 100% refusal compliance on tested benchmarks (n=100, n=28 out-of-sample) with minimal impact on MMLU accuracy (Δ -0.013).
  • Weight-Surgery Method: Utilizes a rank-1 projected, norm-preserving excision of refusal components from 120 residual-write matrices across 40 layers, including hybrid Gated-DeltaNet and full-attention layers, and MoE experts.
  • Vision-Language Model: The 27-block vision tower is fully intact and functional, confirmed by generation tests, allowing the model to process and understand image inputs.
  • Complete Architecture: All 1045 tensors from the parent model are present, including the mtp.* draft head, which is crucial for speculative decoding and often silently dropped in standard transformers conversions.
  • bf16 Safetensors: Provided in full precision, intended as a robust base for fine-tuning.

Intended Use Cases

  • Research into Refusal Mechanisms: Provides a clean model for studying how refusal behaviors are encoded and removed.
  • Red-Teaming: Useful for testing and evaluating the safety and robustness of other AI systems.
  • Unencumbered Fine-Tuning Base: Offers a foundation for developers to fine-tune for specific applications without inherent refusal biases, with the understanding that the user is the primary safety layer.

Finetunes
1 models