Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF

    
        By vramkickedin    
     | 
    
            May 12, 2026 at 7:15 pm        
    
     | 
    
        2 min read

The Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF package delivers an uncensored, performance-enhanced version of Qwen’s latest 27B model in highly accurate compressed formats. This release strips away the original model’s refusal behavior, cutting the refusal rate from 99% to just 4%, then fine-tunes it to beat the censored version on benchmarks. The GGUF quantizations use a custom dual-matrix method that preserves up to 98% of the original model’s accuracy while shrinking file sizes dramatically.

Developer DavidAU who also applied the “Heretic” process to Gemma-4-31B-it-Mystery-Fine-Tune-HERETIC-UNCENSORED-Thinking removes refusals, then used Unsloth for a low-level fine-tune that pushed benchmark scores above the base Qwen 3.6 27B. They built GGUF files with a dual imatrix technique that measures every quant against the full model, ensuring balanced performance for both coding and general text. This project is built for users who need a private, unrestricted assistant that still delivers top-quality results on consumer hardware.

Uncensored model with precision quantization

Key Features

Uncensored: only 4 refusals out of 100 queries.
Fine-tuned to beat original Qwen 3.6 27B scores.
NEO-Code dual imatrix balances code and text.
Q4_K_S quant retains 94% of full accuracy.
256K context window for huge inputs.
Vision support using a separate mmproj file.

This package suits AI enthusiasts with 24GB or larger GPUs who want a private, high-performance model without content filters. Small agencies can deploy the Q4_K_M or Q5 quants locally to handle coding, document analysis, and customer support without cloud costs. Anyone working on sensitive projects or creative writing will benefit from a model that never refuses a legitimate request.

How the quantizations were tested

Every quant file was measured against the full model using rigorous accuracy tests, and the Q8_0 variant includes uncompressed elements for top-tier results. Despite heavy compression, the IQ2_M quant still keeps 83% of the full model’s smarts at a fraction of the size. DavidAU merged two datasets to create the quantization matrix, prioritizing stability so the model handles long conversations and code consistently.