داخل خطة

China's AI landscape shifted dramatically in July 2026, when Zhipu AI founder Tang Jie issued an internal staff letter titled "The Giant Wave Has Arrived," announcing a two-year "Touch High Plan" —a bet-the-company strategy to pour resources into long-horizon AGI research rather than chasing short-term application revenue . This plan's centerpiece: open-sourcing frontier AI models under the most permissive license available, directly confronting Western safety paradigms and U.S. export controls.

The Touch High Vision: Strategic Rationale

The Touch High Plan is a deliberately contrarian strategy announced at a moment of acute market pressure. Zhipu went public in January 2026 , and shortly after the lockup expiry, its stock dropped over 19% . Tang Jie's response was not a defensive pivot but a declaration of aggressive long-term investment .

Key tenets of the Touch High Plan:

• Full return to foundation-model research. The message was clear: no retreat into safe application-layer products, but a doubling-down on the core model .
• Focus on three pillars: long-horizon task capabilities, autonomous agent systems, and self-evolution (self-training)—the capabilities Tang Jie believes necessary to cross the AGI threshold .
• Open-source as a deliberate weapon. Tang Jie explicitly argues that frontier AI should not be "restricted to a small group of developers" .
• "Counterintuitive" strategy. The plan invests in expensive foundational research when competitors are pivoting to monetization .

Tang Jie defines AGI not as a single genius-level model but as "the aggregate of all intelligences across all civilization-level tasks"—a framing that naturally favors broad, open distribution over exclusive control .

GLM-5.2 Under MIT License: The Product Expression

On June 13, 2026, Zhipu released GLM-5.2 under the MIT license—the most permissive open-source license in widespread use . Key specifications:

• 744B total parameters, 40B active (sparse Mixture-of-Experts architecture)
• 1 million-token context window—the longest of any open-weight frontier model
• MIT license: no royalties, no attribution requirement, no regional restrictions. "Organizations can fine-tune, self-host, and build commercial products on top of GLM-5.2 without royalties"
• Ranks as a top-3 model on the Artificial Analysis leaderboard, closely tracking Claude Opus 4.8 and GPT-5.5 on core benchmarks

The timing was critical. Just weeks earlier, the U.S. Commerce Department ordered Anthropic to cut off access to its two most advanced models. Within roughly 18 days, Zhipu answered by releasing a competing model for free . Reuters reported that GLM-5.2 "is at the heart of a growing debate about whether China is finally catching up to the U.S. in the AI race" .

How the Open-Source Stance Challenges Western Safety Paradigms

Tang Jie's position directly confronts the dominant Western safety framework, which holds that frontier model weights are too dangerous for broad release and should be safeguarded by a small number of trusted labs.

The challenge to Western containment logic:

• Capability and containment must scale together. Tang Jie explicitly "rejects the idea that capability should advance faster than containment," arguing that safety research, governance, and model capability should develop in parallel—and that open access enables more eyes on safety, not fewer .
• Rejection of "safety through secrecy." The dominant U.S. paradigm (embodied by OpenAI, Anthropic, and government export controls) holds that the most capable models should remain behind API walls. Tang Jie argues this is both ineffective and anti-competitive—that open inspection produces better safety outcomes than cloistered development .
• Access as a normative good. In a July 12 memo reviewed by Bloomberg, Tang Jie wrote that "frontier AI should remain broadly accessible rather than controlled by select individuals" . This reframes the debate from "open source is risky" to "closed source is an unacceptable concentration of power."

The geopolitical dimension:

The U.S. export-control regime—chip bans on NVIDIA advanced GPUs to China and the Commerce Department's order restricting Anthropic's model access—has created an environment where Chinese open-source releases are a strategic countermove. As Forbes reported, "open source and closed models have become competing strategic narratives in the global competition for AI dominance" . R Street Institute analysis noted that GLM-5.2 performs "within one percentage point of leading American frontier models at a price point roughly one-fifth of the cost" .

Tang Jie's stance weaponizes openness: every time Washington restricts access to U.S. frontier models, Zhipu makes a competitive equivalent freely available, undercutting the premise that restriction can preserve U.S. technological advantage .

Market Performance Context

The Touch High letter came as Zhipu's stock dropped over 19% after IPO lockup expiry . Tang Jie's response was a declaration of aggressive long-term investment, effectively telling shareholders that Zhipu would prioritize AGI research over near-term profitability.

This is a high-stakes bet. The GLM-5.2 MIT release is both a product and a signal: Zhipu can compete at the frontier, and it will do so in public.

Summary Table

Dimension	Tang Jie / Zhipu Position	Western Dominant Paradigm
Model access	MIT license, full open weights	API-only or restricted access
Safety approach	Capability + containment scale together; open inspection	Safety through secrecy, restricted release
AGI strategy	Aggregate of all intelligences; broad distribution	Concentrated in trusted labs
Competitive posture	Open source as strategic counter to export controls	Export controls to preserve advantage
Time horizon	Two-year "Touch High" plan; no short-term monetization	Near-term product revenue + safety alignment

Tang Jie's "Touch High" plan and the MIT-licensed release of GLM-5.2 represent a fundamental challenge to how the world thinks about frontier AI safety and access. Whether openness or restriction proves the safer path is now the defining question of the global AI race.