Qwen3.6-27B abliterated weights arrive on HuggingFace
Iambackup released abliterated BF16 weights for Qwen3.6-27B-AEON-Ultimate-Uncensored on HuggingFace, removing safety layers from the 27-billion-parameter multimodal model.
Iambackup released abliterated BF16 weights for Qwen3.6-27B-AEON-Ultimate-Uncensored on HuggingFace on June 28, stripping safety filters from Alibaba's 27-billion-parameter multimodal model. The checkpoint supports image-text-to-text pipelines and ships in safetensors format with A100 and aarch64 compatibility tags, positioning it for practitioners running inference on both datacenter GPUs and ARM-based consumer hardware. Abliteration rewrites the model's internal representations to bypass refusal behavior without retraining, preserving instruction-following capability while removing content restrictions baked into the original weights.
The BF16 precision keeps memory overhead manageable compared to full FP32 weights—a practical choice for practitioners balancing quality against VRAM constraints on consumer GPUs or cloud instances with limited memory budgets. The 27-billion-parameter count places the model in the mid-tier range for multimodal transformers, large enough for nuanced vision-language tasks but small enough to run on single-GPU setups with 40GB or more of VRAM.
