Tested GLM-5.3 abliterated model: 4x the cost, worse performance than the original

BLUECOW009 · x · 2026-09-03

EnclaveAI tested abliterated-model-large-v2, built on GLM-5.3. Having experienced no refusals from GLM-5.3 itself, they wanted to see if the abliterated version justified 4x the cost. Result: the new model used fewer commands and tokens but failed to match GLM-5.3's performance — making the 4x premium hard to justify.

Original post →

More from Models

Models channel →