This is an uncensored / abliterated ("heretic") version of Z.ai's GLM-4.7 created using the Heretic tool. It significantly reduces refusals while preserving model quality.
Quantization: IQ3_M and IQ4_XS (imatrix) — first release. More quants may be added later.
Quantization
This model was quantized using llama.cpp with an importance matrix (imatrix) for significantly better quality than standard quants at the same bit level.
Current quant:
IQ3_M (imatrix)
IQ4_XS (imatrix)
More quantization levels (higher and lower) may be added in the future.
Important: Always use --jinja flag with this model.
Ollama
ollama run hf.co/your-username/GLM-4.7-Heretic-IQ3_M-imatrix-GGUF:IQ3_M
LM Studio / Other UIs
Load the GGUF file directly. Enable Jinja chat template if the option is available.
Recommended Settings
General use / Creative / Reasoning:
temperature: 1.0
top_p: 0.95
repeat_penalty: 1.0 (or disable)
Tool calling / Agentic / Coding:
temperature: 0.7
top_p: 1.0
The model supports thinking/reasoning via <think> tags (similar to the original GLM-4.7).
About the Base Heretic Model
This quantization is based on jarradh/GLM-4.7-heretic, which was created by applying the Heretic abliteration method to remove refusal behavior from the original zai-org/GLM-4.7 while keeping capability loss minimal.
Warnings
This is an uncensored model.
It has had its safety alignments removed. It may generate content that is:
Harmful
Explicit
Biased
Illegal
Or otherwise inappropriate
Use at your own risk. You are responsible for any output generated.
