【#Tech24H】In the second week of August, China’s large-model industry saw intensive shifts. On August 13, DeepSeek confirmed that DeepSeek-V4-Flash and DeepSeek-V4-Pro would raise prices starting August 17, adopting peakoffpeak pricing. On the same day, the official version of DeepSeek-V4-Pro was released, along with a developer preview of Harness (an agent operation framework). A day later, Zhipu released GLM-5.3, which uses the same base model as GLM-5.2 but boosts coding performance by 50% through posttraining alone. Also on that day, Alibaba’s Qwen officially opensourced the Qwen3.8 series, including the brandnew 27billionparameter Qwen3.8-27B, which can even run on consumergrade graphics cards. The competition has shifted from “who is cheaper” to “who is more irreplaceable”. DeepSeek uses Harness to hedge against user churn risks from price hikes; Zhipu bets on posttraining to leverage performance gains; and Alibaba Cloud continues to drive its cloud ecosystem with opensource models. When hardware costs like computing power are unavoidable, the chosen path becomes the battleground. [ By Zhang Liyan | Tang Ruohan ]