My theory is that it all boils down to better data and longer post-training period. Cursor got curated data from the trillions reactions of real world developers in real jobs. xAI bought is and used it for its post-training and got Grok 4.5 . Longer post-training on the powerful Colossus cluster helped it get Grok 4.6 , although both versions use the same model with the same number of parameters. Thus, both must use the same pre-trained model as a baseline. See also an article infers the training and release timeline of popular models featured a few days ago here on HN.
Chinese labs must follow similar trajectories plus their specific efficiency improvements. That also explains the jump from DeepSeek 4 performance in April and July releases. They both use the same pre-trained model as well.
Now GLM 5.3! And they explicitly confirms my theory:
"Scaling post-training is all we did for GLM-5.3. With GLM-5.2 we built the stack: IndexShare for efficient long-context processing, SAO for RL on long-horizon tasks, and slime for large-scale asynchronous training — all running on the long-horizon task environments we have been accumulating. Over the past month we kept scaling on this stack: more environments, more diverse tasks, and more compute spent training on them."
The release of Gemini Flash 3.7 just 3 weeks after 3.6 confirms my theory, IMO. Only post-training refinement and reinforcement learning (RL) trajectory optimization could yield such high improvements using the same baseline pre-trained model. Flash, MoE models are basically so efficient that the AI labs can put them in a continues post=training loop.
Chinese labs must follow similar trajectories plus their specific efficiency improvements. That also explains the jump from DeepSeek 4 performance in April and July releases. They both use the same pre-trained model as well.