“Today we are releasing GLM-5.3. It uses the same base model as GLM-5.2 — every gain comes from post-training.“