Zhipu AI's 'TouchHigh' Plan Deep Dive: Why the HK-Listed AI Giant Chose to Pursue AGI When Everyone Else Is Chasing Revenue
1. Introduction
On July 11, 2026, Zhipu AI founder Tang Jie published an internal letter titled “The Wave Has Come” (巨浪已来), announcing the launch of the “TouchHigh” (摸高) two-year strategic investment plan — prioritizing AGI breakthroughs over short-term monetization.
The timing is striking.
In the same week, both Zhipu (02513.HK) and MiniMax (0100.HK) — the “twin giants” of HK-listed AI companies — faced their first post-IPO lockup expirations. Zhipu’s stock surged 13.35% on the day of expiration, its market cap briefly touching 906 billion HKD. MiniMax, by contrast, crashed nearly 18%, its market cap falling to just 84.2 billion HKD — down from 410 billion HKD just three months earlier.
While the entire industry races toward monetization, Zhipu has chosen the most “counter-intuitive” path: pushing upward, committing to AGI.
As Tang Jie wrote: “Others ring the bell; we reset to zero. This is not posture — it’s conviction. If the destination is AGI, then short-term gains and industry trends are merely scenery along the journey.”
┌──────────────────────────────────────────────────────────────────┐
│ Zhipu AI 'TouchHigh' Plan Overview │
├──────────────────────────────────────────────────────────────────┤
│ │
│ Strategy: Prioritize AGI over short-term monetization │
│ Duration: Two years (2026-2028) │
│ Four Engines: Long-Horizon Tasks | Autonomous Agents | │
│ Fully Self-Training | Extreme Safety │
│ │
│ Historical Milestones: │
│ 2006 ── Academic search system (single desktop) │
│ 2021 ── GLM-130B (100B parameters, 18 months before ChatGPT) │
│ 2026.01 ── H-Share listing ("reset to zero") │
│ 2026.07 ── "TouchHigh" Plan launched │
│ │
└──────────────────────────────────────────────────────────────────┘
2. Who Is Zhipu: Three Words Behind Two Decades of Methodology
2.1 “Essence, Counter-Intuition, Focus”
Tang Jie’s letter traces Zhipu’s growth using three keywords: Essence, Counter-Intuition, Focus.
2006: The team toiled on an academic search system running on a single desktop. It looked commercially worthless. But the team understood they were “mining the mechanisms of disciplinary evolution” — a question worth ten years to answer.
2021-2022: When “making machines think like humans” was dismissed as moonshot fantasy, Zhipu committed resources to a 100-billion-parameter model — GLM-130B. That was 18 months before ChatGPT exploded.
January 8, 2026: Zhipu listed on the H-Share market. On that day, the company treated it as a fresh start, committing fully to foundational model research.
“Others ring the bell; we reset to zero. This is not posture — it’s conviction.”
2.2 A Demanding Definition of AGI
Tang Jie offered a rigorous definition: “AGI is not the intelligence of a single genius — it is the sum of all human intelligence. It must possess the capability to produce original knowledge at the level of ‘relativity theory.’ That is our sole criterion for determining whether we have truly arrived.”
This definition directly echoes DeepMind’s June 2026 report “From AGI to ASI,” which argued that even if individual model capabilities plateau at human level, as long as compute continues growing, superintelligence may be “squeezed out” through brute force.
3. The Four Engines of the TouchHigh Plan
3.1 Long-Horizon Tasks: From Q&A to Grand Engineering
The first mountain to climb is long-horizon task capability.
Today’s AI excels at instant Q&A. True AGI requires planning and execution spanning weeks, months, or even years.
Zhipu plans to develop a new generation memory architecture that enables models to “learn, do, and remember” across the entire project lifecycle, with the capability to autonomously decompose grand goals (like “design a novel anticancer drug molecule”) into thousands of executable subtasks.
3.2 Autonomous Agent Systems: From Assistant to Digital Employee
The second mountain: building fully autonomous agent systems.
Tang Jie proposed a bold vision — evolving from “one-person company (OPC)” to “no-person company (NPC)”: constructing societies of thousands of specialized agents with distinct “personalities” and “skills,” enabling autonomous debate, collaboration, code review, and resource scheduling.
Notably, Tang Jie pointed out that three challenges once thought to require paradigm shifts — Memory, Continual Learning, and Self-Judgment — are now being solved through existing technology trajectories:
- Long context + RAG approximates memory
- Faster model iteration cycles approximate continual learning
- Frontier models already show signs of self-judgment
3.3 Fully Self-Training: Turning Compute into Evolutionary Fuel
The third and most difficult mountain: Fully Self Training (Fully Self Training).
As high-quality human data approaches exhaustion, Zhipu’s plan is to build a synthetic data factory, using AI self-play to achieve knowledge “from nothing,” and within a safety sandbox, grant systems the ability to rewrite their own code — freeing evolution from the physical limits of human engineers.
“AI training AI is already real — models writing their own code, cleaning and synthesizing their own data, training themselves. This may consume compute, but it saves the most precious resource: human time. In the era of large models, speed is everything. Rapid iteration creates cognitive generational gaps.”
3.4 Extreme Safety Governance: Multi-Billion Resource Commitment
The fourth mountain, and the one Tang Jie emphasized most strongly, is extreme safety governance.
Zhipu plans to invest billions in “Mechanistic Interpretability” — decoding the neural logic behind model decisions, transforming black-box systems into transparent, explainable ones. At the same time, human ethics, social norms, and legal regulations will be written into the model’s value function, with superintelligence alignment research advancing in parallel.
Tang Jie stressed: “The more powerful the capability, the more robust the safety constraints must be. Superintelligence and superalignment must advance together.”
4. Open Ecosystem: Two Sides of the Same Coin
Zhipu’s “TouchHigh” plan is not isolated technical攻坚. Tang Jie emphasized that it’s two sides of the same coin with its open ecosystem strategy:
“One hand reaches upward, challenging the limits of intelligence; the other hand paves the road downward, making the most advanced capabilities as open and accessible as possible.”
On the same day, Zhipu released GLM-5.2 — its most capable open-source model yet, supporting million-level context windows, under the permissive MIT license. Anyone can download, deploy, and commercialize it without restriction.
5. Zhipu vs MiniMax: Divergence of HK’s AI Twin Giants
| Dimension | Zhipu (02513.HK) | MiniMax (0100.HK) |
|---|---|---|
| Unlock size | 6% (manageable) | 60% (financial investors) |
| Unlock day performance | +13.35% | -18% |
| Current market cap | ~731B HKD | ~84.2B HKD |
| Core strategy | TouchHigh, AGI focus | M3 model + API 50% price cut |
| Analyst rating | Institutional buy | JPMorgan downgrade to neutral |
6. From AGI to ASI: The Path Forward
Tang Jie’s letter referenced DeepMind’s “From AGI to ASI” report: even if individual model capabilities plateau at human level, as long as compute continues growing, superintelligence may be “squeezed out.”
DeepMind’s projection: if global deployable AGI instances grow 10x annually, in five years there will be 100 million. These instances, sharing the same underlying brain, with 100x thinking efficiency and zero-cost experience replication, collectively constitute ASI.
Zhipu’s “TouchHigh” plan — long-horizon tasks + autonomous agents + fully self-training + extreme safety — precisely maps to the four critical dimensions of the AGI-to-ASI transition.
7. Conclusion
Zhipu’s “TouchHigh” plan is one of the most strategically significant events in China’s AI industry in 2026:
- No short-term monetization: Choosing the “coldest” path when the industry is hottest
- Four technology engines: Long-horizon tasks, autonomous agents, fully self-training, extreme safety
- Open ecosystem duality: Reaching upward while paving the road downward
- Conviction of HK’s AI first stock: Redefining success with “reset to zero”
As Tang Jie wrote: “The wave has come. Not reaching the summit is failure. This time, the height we must reach belongs to all of humanity.”
This article is based on publicly available reports from The Paper, AI Tech Express, and other media sources.