Shares of Zhipu AI have experienced significant gains for two consecutive trading sessions, driven by speculation that the mysterious high-performance model "Pony Alpha" may be the company's next-generation flagship product. Market rumors suggest that the model, codenamed "Pony Alpha," is actually the upcoming GLM-5 from Zhipu. Amid strong market anticipation, Zhipu AI's stock price surged by a cumulative 60% over two days, reaching a record high.
According to sources familiar with the matter, Zhipu AI is indeed advancing a confidential project, and the mysterious model generating buzz on the OpenRouter platform is highly likely to be the company's soon-to-be-released GLM-5. Boosted by this sentiment, Zhipu AI's market capitalization once exceeded 150 billion HKD, nearly three times its IPO valuation, reflecting investors' high expectations for the iteration of domestic large language model technology.
The catalyst for this market rally was the launch of an anonymous model codenamed "Pony Alpha" on the global model service platform OpenRouter on February 6. Although the model was released in "stealth mode" without disclosing the developer, its demonstrated capabilities in coding, reasoning, and agent workflow were praised by the developer community as "Opus-level" performance. According to Kilo Code, a platform partner, the model is a "specialized evolution of a globally popular open-source model from a certain lab," further fueling market speculation about its origins.
As more technical details emerged, multiple pieces of evidence pointed to Zhipu AI. From self-identification as "I'm GLM" in system prompt tests to tokenizer behavior consistent with GLM-4, and hints from Zhipu's chief scientist about a "new model release around the Spring Festival," market consensus gradually solidified. This stock price movement reflects the capital market's reassessment of domestic large models' global competitiveness in high-end coding and complex logical reasoning.
Key Evidence Points to GLM-5
Regarding the true identity of "Pony Alpha," the technical community uncovered critical clues through multi-dimensional reverse testing. Although the OpenRouter page only indicates that the model has a 200K context window and a maximum output of 131K, during system prompt modification tests, the model directly responded with "I'm GLM." Additionally, Proof-of-Concept (PoC) token tests using specific strings revealed that the model's tokenizer behavior was identical to GLM-4, and its code generation style closely matched that of Zhipu's model family.
Beyond software-level characteristics, architectural details discovered in GitHub code submissions further corroborated GLM-5's technical path. Developers found in a vLLM inference framework pull request that GLM-5 adopts the DeepSeek-V3/V3.2 architecture, incorporating sparse attention mechanisms (DSA) and multi-token prediction (MTP) technology. Based on code analysis, GLM-5 has a total parameter count of 745B, with 78 hidden layers, utilizing a Mixture-of-Experts (MoE) architecture comprising 256 experts, with 8 activated per inference. This architectural change means GLM-5 can maintain high performance while directly benefiting from optimizations in existing inference frameworks like vLLM, significantly lowering deployment barriers.
Programming and Agent Capabilities Stand Out
In practical application tests, the engineering capabilities demonstrated by "Pony Alpha" were the core reason for market attention. According to media evaluations, the model exhibited strong performance in coding, reasoning, and role-playing, being regarded as a Claude Opus-level next-generation flagship foundation model. In front-end development tests, the model could generate a complete radio application with over 500 lines of code from a single prompt, with UI design and interaction logic reaching mature product standards.
More notably was its Agentic Coding capability. In tests replicating complex game projects like "Stardew Valley," the model demonstrated "architectural thinking," autonomously decomposing system requirements, planning project architecture, and continuously programming for over 10 minutes to complete everything from front-end rendering to back-end database setup. Additionally, in refactoring tests for legacy financial system code, the model successfully modernized the code while preserving critical business logic, showing extremely high usability and stability.
Spring Festival AI "Arms Race" Intensifies
The potential release timing of Zhipu's GLM-5 coincides with the密集的 "Spring Festival season" for domestic large model releases. Zhipu AI's chief scientist, Tang Jie, previously hinted in an internal letter that GLM-5 would be released soon, with a timeframe pointing to mid-February 2026. Simultaneously, news has emerged about DeepSeek's new model, MiniMax's planned M2.2 release, and updates to Qwen 3.5 in the near future.
OpenRouter data indicates that over 91% of community users lean towards believing "Pony Alpha" is a beta version of GLM-5. If the model's identity is confirmed as GLM-5, it would not only signify Zhipu's technological breakthrough based on reusing DeepSeek's efficient architecture but also mark a new phase in the competition among domestic foundation models in advanced programming and engineering agent capabilities.
For investors, this technological leap directly translates into a revaluation of Zhipu AI's worth, with the market closely monitoring the company's official release and subsequent commercial deployment.