GLM-5.2: The open coding model that put the critic back into RL

A close look at how Zhipu built GLM-5.2, from DeepSeek Sparse Attention to critic-based PPO, with a live agentic coding demo traced in W&B Weave.