The high-performance model codenamed 'Pony Alpha' is widely speculated to be Zhipu AI's new flagship product, GLM-5, driving its stock price to surge 60% in two days and its market capitalization to exceed 150 billion Hong Kong dollars. Technical testing indicates that the model's architecture and capabilities are highly consistent with the GLM series. Its demonstrated advanced programming and intelligent agent capabilities are regarded as a significant breakthrough for domestically produced large models in key areas. Amidst an intensive period of industry releases, this advancement could reshape the market's assessment of the competitiveness of domestic AI technologies.
Driven by the news that the mysterious high-performance model 'Pony Alpha' is suspected to be Zhipu AI's next-generation flagship product, $KNOWLEDGE ATLAS (02513.HK)$ the stock price of AI-related Hong Kong-listed stocks surged significantly for two consecutive trading days. Market speculation suggests that the model codenamed 'Pony Alpha' is likely GLM-5, which Zhipu AI is about to release. Under strong market expectations, Zhipu AI's stock price rose cumulatively by 60% in two days, hitting a new historical high.
According to informed sources cited by Yicai, Zhipu AI is currently advancing a related confidential project, and the mysterious model that sparked heated discussions on the OpenRouter platform is highly likely to be the company’s upcoming GLM-5. Boosted by this sentiment, Zhipu AI’s market capitalization once exceeded HKD 150 billion, nearly triple its IPO valuation, reflecting investors’ high expectations for the technological iteration of domestic large models.
The trigger for this round of market movements originated from an anonymous model codenamed 'Pony Alpha,' which was launched on February 6 by the global model service platform OpenRouter. Although the model was released in 'stealth mode' without identifying the developer, its capabilities in coding, reasoning, and agent workflows were rated as 'Opus-level' performance by the developer community. According to Kilo Code, a platform partner, the model is described as 'a specialized evolution of one of the most popular open-source models from a certain global lab,' further fueling market speculation about the model’s origins.
As more technical details emerged, multiple pieces of evidence pointed toward Zhipu AI. From self-identification as 'I’m GLM' during system prompt tests, consistent tokenizer behavior with GLM-4, to hints from Zhipu’s chief scientist about 'releasing a new model around the Spring Festival,' market consensus gradually formed. This unusual stock price movement reflects the capital market reassessing the global competitiveness of domestic large models in high-end coding and complex logical reasoning.
Core evidence points to GLM-5.
Regarding the true identity of 'Pony Alpha,' the technical community uncovered key clues through multidimensional reverse testing.
According to media reports, although the OpenRouter page only indicates that the model has a context window of 200K and maximum output of 131K, during system prompt modifications, the model directly responded with 'I’m GLM.' Additionally, Proof-of-Concept (PoC) token tests using specific strings showed that the model’s tokenizer responses were completely consistent with GLM-4, and its code formatting style closely matched Zhipu AI’s model family.
Beyond software-level characteristics, architectural details discovered in GitHub code submissions further corroborated GLM-5’s technical pathway. Developers found in PRs for the vLLM inference framework that GLM-5 adopted the DeepSeek-V3/V3.2 architecture, incorporating sparse attention mechanisms (DSA) and multi-token prediction (MTP) technologies. Based on code analysis, GLM-5 has a total parameter count of 745 billion, 78 hidden layers, and uses a Mixture-of-Experts (MoE) architecture with 256 experts, activating 8 experts per inference. This architectural change means GLM-5 can maintain high performance while benefiting directly from optimizations in existing inference frameworks like vLLM, significantly lowering deployment barriers.
Programming and agent capabilities go viral.
In practical application testing, the engineering capabilities demonstrated by 'Pony Alpha' have been the core factor drawing market attention. According to media evaluations, the model performs strongly in coding, reasoning, and role-playing, and is considered a next-generation flagship foundational model at the level of Claude Opus. In front-end development tests, the model was able to generate a complete broadcasting application containing over 500 lines of code through a single prompt, with both UI design and interaction logic reaching the standards of a mature product.
Even more remarkable is its Agentic Coding capability. During tests replicating complex game projects such as 'Stardew Valley,' the model exhibited an 'architectural mindset,' autonomously breaking down system requirements, planning project architecture, and continuously programming for over 10 minutes to build everything from front-end rendering to back-end databases. Additionally, during tests involving the refactoring of legacy financial system code, the model successfully modernized the code while preserving key business logic, demonstrating extremely high usability and stability.
The AI 'arms race' during the Spring Festival period intensifies.
The potential release timing of Zhipu GLM-5 coincides with the densely packed 'Spring Festival period' for domestic large models. Zhipu AI's Chief Scientist, Tang Jie, previously announced in an internal letter that GLM-5 would be released, with the timeframe pointing to mid-February 2026. Meanwhile, new models from DeepSeek, $MINIMAX-WP (00100.HK)$ including the planned M2.2 and Qwen 3.5, are rumored to receive updates soon.
Data from OpenRouter shows that over 91% of community users believe 'Pony Alpha' to be the test version of GLM-5. If this model is ultimately confirmed to be GLM-5, it not only signifies that Zhipu has achieved a technological breakthrough based on reusing DeepSeek's efficient architecture but also marks a new phase in competition among domestically produced foundational models in advanced programming and engineering agent directions.
For investors, this technological leap directly translates into a reassessment of Zhipu AI’s valuation. The market is closely watching for the company’s official release and subsequent commercial implementation.
Editor/Doris
