Has programming AI changed again? The mysterious model Pony Alpha has become popular, reported to be KNOWLEDGE ATLAS GLM-5, with stock prices soaring 60% in two days!

Driven by the news that the mysterious high-performance model "Pony Alpha" is suspected to be the new flagship product of KNOWLEDGE ATLAS, the stock price of KNOWLEDGE ATLAS on the Hong Kong stock market has surged significantly for two consecutive trading days. The market speculates that this model, codenamed "Pony Alpha," is the GLM-5 that KNOWLEDGE ATLAS is about to release. Under strong market expectations, the stock price of KNOWLEDGE ATLAS has accumulated a rise of 60% over two days, reaching a historic high.

According to First Financial citing informed sources, KNOWLEDGE ATLAS does have related confidential projects currently underway, and the mysterious model that has sparked heated discussions on the OpenRouter platform is likely the GLM-5 that the company is about to release. Boosted by this sentiment, KNOWLEDGE ATLAS's market capitalization once surpassed HKD 150 billion, a figure that is close to three times its IPO market value, indicating investors' extremely high expectations for the iteration of domestic large model technology.

The catalyst for this round of market activity originated from an anonymous model codenamed "Pony Alpha" launched on the global model service platform OpenRouter on February 6. Although this model was released in "stealth mode" without indicating the developer, its capabilities in coding, reasoning, and intelligent agent workflows have been rated as "Opus-level" performance by the developer community. According to Kilo Code, a partner of the platform, this model is a specialized evolutionary version of "the most popular open-source model from a global laboratory," further intensifying market speculation about the model's background.

As more technical details are disclosed, multiple pieces of evidence point towards KNOWLEDGE ATLAS. From the self-identified "I’m GLM" in system prompt testing, to the tokenizer performance consistent with GLM-4, and to the previous hints from KNOWLEDGE ATLAS's chief scientist about "releasing a new model around the Spring Festival," a market consensus is gradually forming. This stock price anomaly reflects that the capital market is 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 has unearthed key clues through multi-dimensional reverse testing.

According to media reports, although the OpenRouter page only shows that the model has a 200K context window and a maximum output of 131K, in the system prompt modification test, the model directly answered "I’m GLM." Additionally, PoC token tests conducted by developers using specific strings showed that the model's response at the tokenizer level is completely consistent with GLM-4, and its code generation formatting style highly matches that of KNOWLEDGE ATLAS models.

In addition to software-level features, architectural details discovered from GitHub code submissions by the open-source community further corroborate the technical path of GLM-5. Developers found in the PR of the vLLM inference framework that GLM-5 adopts the DeepSeek-V3/V3.2 architecture, which includes a sparse attention mechanism (DSA) And multi-token prediction (MTP) technology. According to the code inference, GLM-5 has a total parameter count of 745 billion, with 78 hidden layers, adopting a MoE (Mixture of Experts) architecture with a total of 256 experts, activating 8 at a time during inference. This architectural change means that GLM-5 can directly benefit from optimizations in existing inference frameworks like vLLM while maintaining high performance, significantly lowering deployment thresholds.

Programming and Agentic Capabilities "Going Viral"

In practical application tests, the engineering capabilities demonstrated by "Pony Alpha" are the core reason for market attention. According to media evaluations, this model performs strongly in coding, reasoning, and role-playing, and is considered a next-generation flagship foundational model at the Claude Opus level. In front-end development tests, the model can generate a complete broadcasting application containing over 500 lines of code from a single prompt, with UI design and interaction logic reaching the level of mature products.

Even more remarkable is its Agentic Coding capability. In tests replicating complex game projects like "Stardew Valley," the model exhibited "architectural thinking," able to independently decompose system requirements, plan project architecture, and continuously code for over 10 minutes to complete the setup from front-end rendering to back-end database. Additionally, in tests for refactoring legacy financial system code, the model successfully modernized the code while retaining key business logic, demonstrating extremely high usability and stability.

AI "Arms Race" Upgrade for the Spring Festival

The potential release timing of Zhizhu GLM-5 coincides with the intensive "Spring Festival" release period for domestic large models. Zhizhu AI's Chief Scientist Tang Jie previously hinted in an internal letter that GLM-5 is about to be released, with the time window pointing to mid-February 2026. Meanwhile, news has emerged that DeepSeek's new model, the M2.2 from the MiniMax plan, and Qwen 3.5 will also be updated soon.

OpenRouter data shows that over 91% of users in the community tend to believe that "Pony Alpha" is a beta version of GLM-5. If the model's identity is ultimately confirmed as GLM-5, it not only signifies a technological breakthrough for Zhizhu based on the efficient architecture of DeepSeek but also marks a new stage in the competition of domestic foundational models in high-level programming and engineering intelligence.

For investors, this technological leap directly translates into a re-evaluation of Zhizhu AI's valuation, and the market is closely watching the company's official release and subsequent commercialization developments.