Amid fierce competition in China’s domestic large‑model New Year release window, Zhipu AI officially launched its next‑generation flagship model GLM‑5 late on February 11, 2026, marking another major breakthrough in the company’s foundation‑model efforts. As Zhipu’s first annual flagship product since listing, GLM‑5 features a 744B‑parameter scale and 28.5T of pretraining data, with focused optimizations for coding and agent capabilities, achieving state‑of‑the‑art (SOTA) performance in the open‑source field. The company confirmed that the previously top‑ranked mysterious model "Pony Alpha" on the global model service platform OpenRouter was a disguised test version of GLM‑5. The new model is already available on the chat.z.ai platform; internal evaluations show that its user experience in complex programming scenarios approaches that of top international models such as Claude Opus 4.5, with performance gains of over 20% compared with the previous‑generation GLM‑4. This release not only ignited community enthusiasm but also propelled Zhipu’s Hong Kong‑listed shares to surge nearly 60% in two days, highlighting the accelerating catch‑up of domestic AI through independent innovation. The arrival of GLM‑5 may reshape the open‑source large‑model ecosystem; below we analyze the technical strengths and market potential of this "domestic dark horse."
Context of the GLM‑5 Release: Zhipu AI’s Strategic Upgrade and New Year Competition
Zhipu AI was founded in 2019, incubated by the Knowledge Engineering Lab at Tsinghua University, and is a leading Chinese AI startup. The company went public on the Hong Kong Stock Exchange on January 8, 2026, becoming another AI unicorn after SenseTime. On the day of listing, Chief Scientist Tang Jie, a professor at Tsinghua University, announced in an internal letter that GLM‑5 would be released soon and declared a full return to foundational model research starting in 2026, establishing the frontier innovation unit X‑Lab to focus on architecture, learning paradigms, and continual evolution. The GLM series has been Zhipu’s core product line since GLM‑3 in 2023 and has iterated to GLM‑4.7, covering language, vision, and multimodal capabilities. The launch of GLM‑5 comes at the peak of China’s New Year period for domestic large models: following DeepSeek V4 and ByteDance’s Seedance 2.0, GLM‑5’s appearance further intensifies competition.
The timing of the release was clever: February 11 fell just before the Lunar New Year, allowing Zhipu to capture developer attention. As early as late January, the "Pony Alpha" model on the OpenRouter platform went viral for being free, offering a ~200K context window and strong coding ability, and was suspected by the community to be a disguised GLM‑5. Zhipu later confirmed this identity to the media, calling it a testing strategy to gather global feedback. The "mystery marketing" successfully propelled GLM‑5 to trending status, with discussions on X calling it "open‑source SOTA, national pride."
Technical Highlights: 744B Parameters and Sparse Attention Similar to DeepSeek
The core of GLM‑5 lies in its scale and architectural optimizations. The model’s total parameter count reaches 744B (744 billion), activation parameters increase from 32B in the previous generation to 40B, and the pretraining data volume reaches 28.5T. Zhipu adopted the technical route of DeepSeek V3/V3.2, integrating a sparse attention mechanism (DSA) and multi‑Token prediction (MTP), significantly improving long‑sequence handling and efficiency. The DSA mechanism allows the model to maintain high performance while reducing computational complexity, making it suitable for complex systems engineering and long‑horizon agent tasks.
On coding, GLM‑5 achieves SOTA in the open‑source domain. Internal evaluations indicate more than a 20% improvement over GLM‑4 in frontend, backend, and long‑horizon tasks. On the SWE‑Bench Verified leaderboard, GLM‑5 ranks in the top tier and is comparable to Claude Opus 4.5. Developers report that in real‑world programming scenarios, the user experience approaches top international levels, handling complex codebases, bug fixes, and architecture design. For example, in frontend development GLM‑5 can generate responsive UI code; for backend tasks it supports database optimization and API integration.
Agent capabilities are another major highlight. GLM‑5 excels at long‑horizon tasks such as multi‑turn automated workflows, can simulate multi‑agent collaboration, and handle the full chain from planning to execution. Compared with GLM‑4, the new model is more robust in tool invocation and environment adaptation, making it suitable for game development, automated testing, and similar scenarios. Zhipu emphasizes that GLM‑5’s open‑source strategy will spur community innovation and has made code available on GitHub.
Additionally, GLM‑5 supports a 128K context window and a maximum output length of 8192, strengthening multimodal capabilities. The company also launched companion services such as the GLM Coding Plan (20 CNY/month), offering dedicated MCP (model‑call priority) and tool compatibility. This positions GLM‑5 not just as a model but as part of a developer ecosystem.
Comparison with the Previous Generation: Leap from GLM‑4 to GLM‑5
The GLM‑4 series (including GLM‑4.7) was released in 2025 and led in language understanding and vision tasks. GLM‑5, however, doubles the scale from 355B to 744B parameters and increases data from 10T+ to 28.5T. Performance gains are significant: on leaderboard aggregators such as Artificial Analysis Intelligence and LM Arena, GLM‑5 ranks first among open‑source models and first among domestic models. Compared with contemporary products like DeepSeek V4 and Qwen 3.5, GLM‑5 stands out on programming SOTA metrics, and its API revenue surpasses that of other domestic models combined.
Community tests show GLM‑5 is more robust in complex task execution and creative writing. For example, on the SWE‑Bench benchmark GLM‑5 scored 77.8%, second only to Claude Opus 4.5’s 80.9%. This is attributed to the introduction of the DSA mechanism, which lowers training costs while improving long‑sequence reasoning.
Market Impact: Stock Surge and Reshaping the Domestic AI Ecosystem
After the GLM‑5 release, Zhipu’s Hong Kong‑listed stock jumped nearly 60% in two days, with market value approaching the low hundreds of billions. This reflects investor confidence in domestic AI: GLM‑5’s open‑source SOTA position could attract more developers and challenge international incumbents. Zhipu says the model will empower sectors such as education, healthcare, and entertainment and help drive progress toward AGI (artificial general intelligence).
In the New Year release window, GLM‑5, DeepSeek V4, and Minimax 2.5 form a competitive "top three." Analysts believe this will accelerate domestic large models’ shift from following to leading and narrow the gap with models like GPT‑5.
Outlook: Zhipu’s Path to AGI and the Challenges Ahead
The release of GLM‑5 marks the start of Zhipu’s renewed focus on foundational research. The company plans to deepen X‑Lab in 2026 to explore new architectures and continual learning. However, challenges remain: while parameter scale is large, open‑sourcing requires balancing intellectual property; and although agent capabilities are strong, safety alignment still needs reinforcement.
Overall, GLM‑5 represents a new high for domestic AI. It is more than a model upgrade—it is an ecosystem revolution. Are you ready to experience this "open‑source leader"? Zhipu’s next steps may help define the future of Chinese AI.
▶️ Enjoy this article? Content creation is hard!
Follow our YouTube channel:
AI Rosetta
Your likes, subscriptions, and shares are our greatest support! 🙏
For more AI news visit: https://www.airosetta.com