Qwen 3.6 40B abliterated fine-tune lands on HuggingFace with Claude and Deckard datasets

Black6spdZ released an abliterated 40-billion-parameter fine-tune of Qwen 3.6 on HuggingFace on May 18, 2026, blending Claude 4.6 Opus, Deckard, and Heretic training datasets in quantized GGUF format. The model targets practitioners running multimodal workflows locally without safety enforcement, with imatrix optimization to fit the checkpoint onto consumer GPUs with 24GB VRAM or less.

The release sits in the image-text-to-text pipeline category, positioning it for workflows that combine vision and language tasks. The model card tags the checkpoint as "uncensored," "abliterated," and "multi-stage tuned," signaling removal of alignment guardrails through iterative fine-tuning passes. Training relied on Unsloth's memory-efficient fine-tuning library and standard transformers infrastructure.

The naming convention references three distinct training sources: Claude 4.6 Opus data, the Deckard dataset, and Heretic fine-tuning stages. The "NEO-CODE" and "Di-IMatrix" tags point to the quantization and optimization passes applied before GGUF export. Imatrix quantization uses importance matrices to preserve model quality at lower bit depths, a standard technique for fitting 40B models onto consumer hardware. The GGUF format, maintained by the llama.cpp project, remains the de facto standard for CPU and GPU inference outside Python runtimes.

Qwen 3.6, the base model, is Alibaba's open-weight multimodal series released earlier this year. The 40B parameter count places it in the mid-range for local deployment—larger than 7B and 13B consumer favorites but still runnable on high-end consumer hardware with aggressive quantization. At publication, the weights showed zero downloads and zero likes.