1. GLM-5.3, while maintaining its 744 billion parameters, primarily improves performance through post-training expansion, achieving significant progress in benchmarks such as encoding, long-task execution, cost efficiency, and cybersecurity. It reportedly surpasses DeepSeek V4 Pro (with 1.6 trillion parameters) in multiple tests.
2. Zhipu's core competitive advantage isn't just its models. With over 1 million active users on ZCode, it has formed a data flywheel of "user usage → real-world task data → feedback for model training," helping to maintain GLM's cost-effectiveness advantage in coding scenarios. The idle-time task mechanism helps improve the utilization of inference infrastructure.
3. Industry pricing shows signs of normalization. Zhipu's encoding package discounts are narrowing, and the DeepSeek V4 Pro API price is significantly higher than V4 Flash, indicating that the improved capabilities of cutting-edge models are translating into pricing power.
4. Goldman Sachs maintains a "Neutral" rating on Zhipu with a target price of HK$1610. They believe Zhipu still has significant long-term commercialization potential, and the current ARR may underestimate real demand. Future revenue potential can be unlocked through open weights and commercial licensing.
5. The risks lie in its still high valuation, and the challenges it faces including intensified competition, high R&D investment, continued losses, and geopolitical risks. (The above views are from Goldman Sachs' report dated August 17.)
