> this is just GLM 5.2 with post-training magic
Isn't post-training turning out to be the most important part?
It basically has been ever since they started using RLVR for reasoning (esp. coding & math), with the DeepSeek-R1 paper being what let the cat out of the bag.
The Gemini 3.7 Flash model released yesterday, and all the 3.x Flash models, are still based on the Gemini 3 pre-training run from January 2025 !!