I just came across the following post, where a user found some confusing divergence results between Q4 quants of the original and QAT models with a Q8/unquantized reference of the original model. https://www.reddit.com/r/LocalLLaMA/comments/1tyxu55/gemma_4_31b_qat_q4_vs_standard_q4_top1_kld/ From there I understood that after the retraining of Gemma 4 31B QAT, this could be considered as a different model to Gemma 4 31B original. Therefore, it is not useful to test the divergence of Gemma ...