GLM 5.2

The latest burst of optimism in China’s AI sector has come from Zai founder Tang Jie, who believes a Mythos-class model could emerge before the first quarter of 2027. His view landed just as the company’s GLM 5.2 drew attention for its strong benchmark performance. The timing matters because Mythos-class systems are associated with advanced cyber capabilities. They have been described as able to break software defenses and identify tens of thousands of previously undiscovered bugs, a prospect that has already shaped policy choices in Washington. The United States has already responded with tighter controls on the AI race. After restricting exports of advanced AI chips to China, the White House also moved to limit access to Anthropic’s Mythos 5 and Fable 5 for all foreign nationals. According to reports cited in the broader discussion, U.S. officials fear groups linked to China could jailbreak those models and use them for cyberattacks. That concern is one reason a domestically built Chinese alternative with similar capabilities is being watched so closely. Why GLM 5.2 shifted the conversation GLM 5.2 gave Tang Jie’s comments more weight because the model is now the leading open weights system on the Artificial Analysis Intelligence Index with a score of 51. Open weights means the model’s learned parameters are public, allowing anyone to download and run it locally. The ranking places GLM 5.2 ahead of other Chinese systems such as Kimi K2.6, and above Google’s Gemini 3.5 Flash at 50 points and Anthropic’s Claude Sonnet 4.6 at 47. Even so, it still trails the top American models, with Fable 5 at 60 and OpenAI’s GPT 5.5 at 55. Its leap is especially notable because the model is the same size as GLM-5.1, at 744 billion total parameters and 40 billion active parameters. Despite that unchanged scale, it improved by 11 points on the index. Huawei chips and a lower training bill Another reason GLM 5.2 drew attention is its hardware path. The model was trained entirely on Huawei Ascend chips, without Nvidia hardware, which makes it an important case for China’s push to reduce reliance on U.S. technology. That detail matters because U.S. export limits on advanced chips have long been seen as one of the biggest constraints on China’s AI ambitions. If competitive models can be trained on domestic alternatives, the pressure from those limits becomes less effective over time. Emad Mostaque, founder of Stability AI, estimated that the total training cost may have been around $25 million. He said that figure is far lower than the sums likely needed by Anthropic or OpenAI to train their own models. The combination of lower cost and local chips changes the competitive picture. It is no longer only a contest over raw capability, but also over who can build frontier models with a more controlled supply chain and a more manageable budget. The wider rivalry is still moving The exchange that followed on X also showed how quickly expectations are shifting. A user suggested that Zai could build a Mythos-level model after GLM 5.2, pointing to a possible window between Nov-Des ’26. Elon Musk replied that such a milestone would more likely arrive in the first quarter of 2027, but Tang Jie pushed back and said the timeline would not take that long. The disagreement added fuel to a rivalry that is already defined by speed, access, and technical scale. There is still a cautionary note in the background. In the past, U.S. AI companies have accused several Chinese firms of distillation, a technique that trains a smaller model using the output of a larger one so it can imitate its behavior. Even with that dispute unresolved, the broader trajectory is clear. China may be moving closer to a Mythos-class system faster than some expected, while companies such as Anthropic and OpenAI are also likely to keep improving their models during the same period. Source: www.indiatoday.in

China’s AI Race Tightens, Zai Says Mythos-Level Model Could Arrive Before 2027