GLM 4.7 Flash Benchmark

4 abliteration techniques were compared on GLM-4.7-Flash using the same tools and procedures.

The techniques compared were Heretic, HauhauCS Aggressive, Abliterlix, and Huiui, and the comparison went beyond simple benchmarks to include safety evaluation, weight analysis, KL divergence, and chain-of-thought forensics.

In particular, since GLM-4.7-Flash uses a Mixture of Experts architecture with 64 routed experts per layer, the author noted that how abliteration works could differ from the patterns seen in the Qwen family.

Building on the existing claim that HauhauCS's tool is a plagiarized fork of Heretic, the author also analyzed what cost the added third-party techniques leave on model performance.

The results were published as a benchmark collection and a safetensors analysis page on HuggingFace.