Zhipu Responds to Criticism of 'Post-Training Only': Next-Generation Foundational Model GLM-6.0 Targets 'Self-Evolution'

During the 2026 semi-annual performance communication meeting held on the evening of August 31, an analyst raised concerns from the market regarding Zhipu's 'post-training only' approach and inquired about the company's recent restraint in parameter scale expansion. Zhipu's founder Tang Jie stated that the next-generation model will still expand the foundational scale while implementing controls on activation parameters to avoid inference speed degradation and cost increases. On the computing power front, Zhipu disclosed that it has achieved large-scale inference using domestically produced chips at the 100,000-level, with per-token inference costs reduced by 80% compared to the beginning of the year. GLM 5.3 Flash is Zhipu's first model to fully rely on domestic chip clusters for service under ultra-large-scale real-world traffic. The company claims that compared to the initial baseline of the same hardware, the model's end-to-end service performance has improved threefold. Tang Jie summarized one direction for the future GLM-6.0 as self-evolution: 'This is a significant challenge. The biggest issue is whether the model can determine for itself when to stop and when to correct itself. This is the biggest problem, rather than simply making the model larger. In future models, this will be a key area of research.' (Everyday Economic News)