Huihui GLM 5.2 Abliterated

Huihui-GLM-5.2-abliterated-GGUF is a text-generation model based on zai-org/GLM-5.2, with safety refusals removed via abliteration (the technique used in remove-refusals-with-transformers). The abliteration directly modifies GGUF files from unsloth/GLM-5.2-GGUF by converting some weights originally quantized at Q5_K and Q6_K to Q8_0 (for the first 12 layers and all expert modules) and some to MXFP4 (UD-Q2_K_XL → UD-Q2_K_MXFP4). The resulting model size increase is modest — for example, the UD-IQ1_M variant is approximately 3 GB larger than the original.

Key Strengths

• Uncensored operation: The model’s refusal mechanism has been largely disabled, allowing it to generate responses that standard models would typically reject.
• Direct ablation on GGUF: The modification works at the quantized weight level, preserving most of the original model’s architecture and core capabilities.

Usage Considerations

This model is intended for research, testing, or controlled environments. Its reduced safety filtering creates several risks:

• Outputs may include sensitive, controversial, or inappropriate content.
• Not suitable for public settings, underage users, or high-security applications.
• Users bear full legal and ethical responsibility for generated content.
• Real-time monitoring and manual review are strongly recommended.

No benchmark performance numbers are provided in the model card. The technique is described as a “crude, proof-of-concept” implementation and may not be compatible with all model architectures when applied to other bases.