Tang Jie doesn’t mince words. The founder of Zhipu AI, one of China’s standout artificial intelligence companies, believes frontier models should remain open to all. He laid out that position in a recent internal memo. Safety, he contends, stems from widespread involvement, shared knowledge and collective scrutiny. Not from walls that keep technology locked away from most eyes.
His stance lands at a delicate moment. Beijing is reportedly considering restrictions on overseas access to the country’s most advanced open models. The timing feels pointed. Zhipu itself just released GLM-5.2 under a permissive open-source license. Anyone can download it, modify it and even build commercial products on top. The model supports a full 1 million token context window. It delivers strong results on coding, agentic tasks and long-horizon reasoning.
According to The Next Web, Tang’s memo inverts conventional security thinking. Real protection comes when many independent researchers examine a system. Flaws surface faster that way. A small team working in isolation misses things. The argument echoes concerns raised by cybersecurity professionals elsewhere. When U.S. authorities restricted a powerful model, more than 100 experts signed an open letter warning that defenders suffered most.
But the counterargument carries weight too. Once weights are public, they cannot be retracted. Bad actors gain the same access as researchers. Safeguards built into the model can be removed with modest effort. The debate has no easy resolution. Both sides highlight genuine dangers. Empirical evidence that settles the question remains elusive.
Zhipu’s move with GLM-5.2 is more than symbolic. The model lands within a percentage point of Anthropic’s Opus 4.8 on a prominent agentic benchmark. It does so at roughly one-fifth the cost. CNBC reported the performance numbers and noted the buzz it created in Silicon Valley circles. Gabe Pereyra, co-founder of Harvey, told the publication he has been surprised by how quickly open-source offerings have caught up. GLM-5.2 represents the first time an open model feels truly competitive with certain closed frontier systems.
The company’s stock reacted sharply. Shares of the Hong Kong-listed entity surged after the release. Zhipu, which rebranded internationally as Z.ai, went public earlier in 2026. Its initial public offering drew strong investor interest. The firm has raised billions in prior rounds. That capital has fueled rapid progress despite export controls on advanced chips.
Tang, who also holds a professorship at Tsinghua University, has steered the organization since its founding in 2019 as a university spin-off. The team started with knowledge-graph expertise before pivoting hard into large language models. Early versions of the GLM series established a reputation for efficiency. Later iterations pushed into multimodal capabilities and sophisticated agent behavior.
In his internal letter titled “The Great Wave Has Arrived,” reviewed by multiple outlets, Tang outlined the company’s direction. Zhipu will pour resources into long-horizon reasoning, autonomous agents and self-improving systems. The firm plans to keep pushing model boundaries while maintaining its open approach. Yahoo Finance covered the memo’s key points, quoting Tang on the need for broad participation and public oversight instead of access limits.
The Chinese government’s reported deliberations add tension. Reuters first broke the story that officials are weighing limits on foreign downloads of top open models. Such a policy would mark a reversal. Open-source releases have been a strategic strength for Chinese labs. They allowed rapid global adoption. Developers worldwide now rely on models from Zhipu, Alibaba’s Qwen team and DeepSeek. Chinese open-weight systems account for more than 45 percent of traffic on some inference routing platforms.
Yet national security officials worry about uncontrolled proliferation. Advanced capabilities in the wrong hands could enable bioweapons research, large-scale cyberattacks or other threats. The same models that power helpful agents could be stripped of safety training and turned to harmful ends. This mirrors debates in Washington, though the contours differ.
Tang’s position puts him at odds with potential regulatory shifts in his own country. He argues that closing off access would slow collective progress on safety. More eyes improve defense. Fewer eyes create blind spots. His view aligns with a community of researchers who see openness as a net positive for security.
Performance data backs Zhipu’s momentum. On several coding and agent benchmarks, GLM-5.2 ranks among the top three globally alongside closed models from Anthropic and OpenAI. It achieved a score of 81 on Terminal-Bench, a significant jump from the prior version. The model handles complex software engineering tasks with greater reliability. Its long context window allows it to process massive codebases or lengthy documents without losing coherence.
Cost advantages matter just as much. Inference pricing for GLM-5.2 sits far below comparable U.S. offerings. That accessibility has driven adoption among startups, researchers and enterprises seeking capable AI without massive budgets. Reuters highlighted how the release narrowed the gap with Western labs at a time when some U.S. companies faced their own setbacks.
Investors have taken notice. Zhipu’s market value climbed dramatically in the months following its listing. The stock rose more than 20-fold in a six-month stretch at one point. Observers see the company as well positioned to fill demand left by restricted American models. But that bet depends on continued openness.
Industry conversations reflect the stakes. In a wide-ranging discussion among leaders from Zhipu, Moonshot, Alibaba and Tencent, Tang cautioned against over-optimism. He noted that while open-source releases create excitement, the gap with closed American systems may still be widening in certain respects. The comments, captured by ChinaTalk, underscore a candid assessment amid public hype.
Still, Zhipu’s track record shows determination. The firm iterated quickly through multiple GLM generations. It optimized aggressively for domestic hardware when Nvidia chips became harder to obtain. Compatibility with Huawei Ascend processors and other local accelerators became a priority. Those engineering choices allowed continued scaling.
Tang has emphasized practical utility. The company’s agent products, including AutoGLM and GLM-PC, demonstrate real-world capabilities. They can perform multi-step operations across applications and even control browsers autonomously. Open-sourcing base models for these agents invites the community to build upon them.
The broader context includes geopolitical friction. U.S. export controls on advanced semiconductors aimed to slow Chinese AI development. Beijing responded with its own measures and heavy investment in domestic supply chains. The result is a bifurcated technology world. Yet open models have served as a bridge, letting capabilities spread despite hardware barriers.
If China imposes new limits, that bridge could crumble. Global developers might lose access to the most competitive open offerings. The irony is thick. Chinese labs helped popularize affordable, high-performing models. Now their own government may restrict the very approach that fueled that success.
Tang cannot dictate policy. His memo reads as an attempt to influence the conversation from within. He presents openness not as naive idealism but as strategic realism. In his telling, the alternative creates greater vulnerabilities by concentrating knowledge and reducing scrutiny.
Whether Beijing listens remains uncertain. Closed-door meetings involving Zhipu, Alibaba and ByteDance have already taken place. The outcome will shape the next chapter of AI development. It will influence who gets to work with the most powerful systems and on what terms.
For now, GLM-5.2 stands as a statement. It is available under the MIT license with no restrictions on commercial use or user types. The company’s API pricing makes it accessible. Its benchmarks position it near the frontier. And its founder continues to make the case that keeping such technology open serves everyone, including those tasked with keeping it safe.
The coming months will test that proposition. Regulators in both Washington and Beijing are reexamining their approaches to frontier AI. The decisions they reach could determine whether the technology remains a shared resource or becomes increasingly siloed. Tang Jie has placed his bet. The world is watching to see if others follow.
WebProNews is an iEntry Publication