The domestic large model sector saw multiple breakthroughs during the 2026 Spring Festival. After DeepSeek became a phenomenon, Zhipu AI's new generation large model GLM-5 also unveiled its mystery. This move directly triggered the capital market, with Zhipu stock rising by 200% recently, and the total market value reaching 150 billion HKD, three times that of its IPO.

Identity Revealed: The Mysterious Model "Pony Alpha" is GLM-5

Recently, a model named "Pony Alpha" appeared on the global model service platform OpenRouter. Due to its code writing capability approaching that of Claude Opus, it sparked a global discussion.

  • Confirmation of Identity: The system prompt of this model revealed its GLM identity.

  • "Fingerprint" Identification: Netizens identified its归属 through verifying the unique logical bugs of the GLM family (such as getting specific abnormal answers when inputting "heat vegetable oil in the pan"), almost certainly confirming its origin.

Core Black Technology: Reusing DeepSeek Architecture, Doubling Parameters

GLM-5 chose the same sparsity attention architecture (DSA) as DeepSeek-V3, which is seen as an extremely cost-effective evolution strategy.

  • Scale Leap: The total number of parameters reached 745B, twice that of the previous GLM-4.7.

  • Computational Efficiency: It has 256 experts, with 8 activated each time (about 44B activated parameters), and a sparsity rate of only 5.9%.

  • Long Text and Multimodal: It supports a maximum context window of 202K tokens. At the same time, in response to the market demand in 2026, GLM-5 enhanced multimodal capabilities such as video understanding, compensating for the shortcomings of DeepSeek's pure text architecture previously.

Industry Impact: Deployment Threshold Further Reduced

Due to the use of DSA architecture, GLM-5 can directly reuse existing optimization schemes from mainstream inference frameworks such as vLLM and SGLang. This means that enterprise users will significantly reduce technical barriers and computing costs when deploying this model.

In the wave of domestic AI "stealing" overseas large models, Zhipu once again proved its top-level capabilities in model performance and engineering implementation with the strong performance of GLM-5.