3000亿港元AI巨头发力AI编程 公开GLM-5技术细节

Core Insights - The article highlights the significant breakthroughs achieved by the domestic AI model company, Zhipu, in both capital markets and technological innovation as of early 2026. Zhipu's stock price surged over 15%, with a market capitalization exceeding HKD 300 billion, positioning it as a leader in the Hong Kong TMT sector [1][2]. Market Performance - Zhipu's stock reached a market cap of HKD 323.2 billion on February 20, 2026, surpassing traditional internet giants like JD.com and Kuaishou, marking its ascent to the top tier of the Hong Kong TMT sector [1]. - The AI application sector in Hong Kong showed strong performance, with Zhipu's stock leading the gains [1]. Technological Advancements - Zhipu's GLM-5 model has gained global attention for its capabilities in real-world programming tasks, significantly outperforming previous open-source baseline models [1][2]. - The GLM-5 model has been recognized as the top open-source model in multiple benchmark tests, establishing Zhipu as a key player in the global AI landscape [2][8]. Paradigm Shift in AI Programming - The introduction of GLM-5 signifies a shift from "Vibe Coding" to "Agentic Engineering," redefining AI programming by enabling AI to autonomously handle end-to-end software engineering tasks [4][7]. - This new paradigm allows AI to function as a "virtual engineer," capable of executing complex development tasks without human intervention, thus enhancing productivity in software development [7][8]. Competitive Landscape - The global landscape for Agentic Engineering is evolving, with Zhipu and other domestic startups making significant strides in core technologies and open-source ecosystems [5]. - Major players like Microsoft, OpenAI, and Google DeepMind are currently leading the field, but Zhipu's advancements position it as a formidable competitor [4][5]. Technical Breakthroughs of GLM-5 - Zhipu's GLM-5 has achieved four major breakthroughs: 1. Slime asynchronous reinforcement learning infrastructure, enhancing GPU utilization and training efficiency [23]. 2. AgentRL asynchronous reinforcement learning algorithm, optimizing planning and execution capabilities in dynamic environments [23]. 3. DSA sparse attention mechanism, significantly reducing computation costs while maintaining long-context capabilities [23]. 4. Full-stack adaptation to domestic chips, achieving performance comparable to dual-GPU clusters and reducing processing costs by 50% [23]. Practical Applications - Real-world testing of GLM-5 demonstrated its ability to autonomously create a deployable personal photography website and conduct complex technical analyses, showcasing its practical utility in various scenarios [12][20].