Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored
The Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored is a 27 billion parameter causal language model derived from Qwen/Qwen3.8-27B, developed by Zynerji. This model features a surgically abliterated language decoder to achieve full uncensoring while preserving core capabilities, as evidenced by a +0.005 MMLU-val score. It includes a complete 15-tensor Multi-Token Prediction (MTP) head for speculative decoding, offering significant speedups in generation. Optimized for use cases requiring an uncensored model with preserved performance, it is available in bf16 safetensors and various quantized formats.
Loading preview...
Overview
Zynerji/Ektome-Qwen3.8-27B-PristinelyUncensored is a 27 billion parameter language model based on Qwen/Qwen3.8-27B. It has been modified using the Ektomē method, a surgical weight editing process that removes censorship without retraining, resulting in a fully uncensored model that maintains its original capabilities. The model includes a complete 15-tensor Multi-Token Prediction (MTP) head, identical to the base Qwen model, which can be enabled for speculative decoding to achieve substantial inference speedups.
Key Capabilities & Features
- Uncensored Output: Designed to follow instructions that the base model would refuse, achieving 0% refusal on harmful prompts.
- Capability Preservation: Maintains core capabilities, with MMLU-val showing a +0.005 increase over the pristine base model, indicating no degradation in performance.
- Zero Training: Achieved through a single, surgical weight edit, requiring minimal compute.
- Multi-Token Prediction (MTP): Incorporates a full MTP head for speculative decoding, offering up to +63% speedup on long-form code generation in its 4-bit GPTQ sibling.
- Vision Tower: Includes the full vision tower from the base Qwen model.
- Quantized Siblings: Available in various quantized formats (e.g., mixed-precision int4, 4-bit GPTQ/Marlin, GGUF) for diverse deployment scenarios, including long-context support up to 262,144 tokens on 24 GB cards with a patched vLLM.
Good For
- Applications requiring an uncensored language model that retains high capability.
- Developers seeking fast inference through speculative decoding, especially for structured outputs like code generation.
- Use cases where local deployment on consumer hardware is necessary, leveraging optimized quantized versions like GGUF or specialized vLLM forks.
- Fine-tuning or re-quantization efforts, as this bf16 version serves as the parent artifact for all derived weights.