@adlrocha - Base Models Stopped Being the Bottleneck

Last week I closed the Kimi K3 post promising you that I’d do the same exercise of explaining the improvements of the newly released Qwen3.8 and GLM5.3 models in plain English. It turns out that as I was writing this post Qwen and GLM decided to also release their new Flash models (marking that day in the calendar AI independence day), so I guess this is going to become a three post series after all.One of the things that I really liked about GLM5.3 and Qwen3.6, and that is worth analysing, is that the models don’t introduce any architectural change.What GLM5.3 isGLM-5.3 shipped on August 14 on the same base as GLM-5.2 (with the same size and number of parameters) but with an additional month of post-training. Z.ai summarises the release as follows: “Scaling post-training is all we did for GLM-5.3.” That was enough, a month of training later, and the same underlying brain made it into the top ranking of CyberGym and GDPval.Through that month of training, the model managed to get close…