Zhipu CEO Tang Jie’s Internal Letter: “The Great Wave Has Arrived”
Image: Official release of Zhipu’s internal letter titled “The Great Wave Has Arrived”
🌊 A Milestone in AI Leadership
On July 11, 2026, Zhipu AI Founder and CEO Tang Jie issued a landmark internal letter titled “The Great Wave Has Arrived” — marking not only a corporate inflection point but a defining moment for China’s generative AI ecosystem.
In just six months since its H-share listing, Zhipu achieved extraordinary growth:
– Market capitalization surged 10× from IPO levels;
– Joined the “Trillion HKD Club” in June 2026 — valued at ~HK$1.02 trillion, nearly 3× Baidu’s market cap and surpassing Xiaomi;
– Maintained stable share price post-first-share-lockup on July 8 — signaling strong investor confidence.
This trajectory reflects what the industry now calls “the sexiest story in large language models”: disciplined technical conviction yielding both market validation and commercial traction.
🔍 The Strategic Bet: Coding as the Catalyst
Zhipu’s ascent traces back to a pivotal 2025 decision: reallocating R&D resources toward coding capability — a move grounded in deep technical foresight.
At a public forum, Professor Tang Jie (also Chair of Tsinghua University’s Department of Computer Science) articulated the rationale:
“The emergence of DeepSeek R1 signals the end of the Chat paradigm. In the post-DeepSeek era, we’re betting on two foundational capabilities: Coding and Reasoning — abilities that enable true symbiosis with intelligent agents.”
That bet paid off:
– GLM-4.5 (released July 2025) laid the groundwork;
– GLM-5.2, open-sourced in June 2026, now ranks among the world’s top-tier AI coding models, matching or exceeding benchmarks of Claude Opus 4.8 and GPT-5.5 across core metrics;
– Zhipu’s MaaS platform reached ¥1.7 billion ARR (as of March 2026), a 60× YoY increase.
🧭 The Next Frontier: Three Pillars of AGI Advancement
Tang Jie’s letter pivots decisively beyond coding — outlining Zhipu’s next strategic horizon centered on three interlocking pillars:
✅ Long-Horizon Task Capability
Enabling models to execute complex, multi-week/multi-month workflows — e.g., autonomously auditing software for zero-day vulnerabilities by emulating expert security analyst cognition — not just answering queries, but orchestrating outcomes.
✅ Fully Autonomous Agent Systems
Moving from “AI assistants” to self-sustaining, collaborative agent ecosystems — equipped with persistent memory, continual learning, and self-judgment — evolving toward “NPC companies” that operate 7×24 without human intervention.
✅ Self-Evolving Intelligence
Advancing AI training AI (AI²): models that write code, synthesize high-fidelity training data, and iteratively retrain themselves — leveraging massive compute (e.g., million-GPU clusters) to compress R&D cycles and widen cognitive lead times.
“When AGI instances scale tenfold annually, one billion instances within five years could collectively manifest ASI — not via algorithmic leaps alone, but through sheer scale, shared cognition, and zero-cost knowledge replication.”
— Citing Google DeepMind’s From AGI to ASI report
🚀 The “Touch High” Strategic Initiative
To pursue these frontiers, Zhipu launched Project Touch High: a multi-year, non-commercial-first investment in AGI’s physical and algorithmic frontiers.
Four core engines drive the initiative:
| Engine | Focus Area | Key Objective |
|---|---|---|
| 1. Long-Horizon Task | Memory & Planning Architecture | End-to-end project lifecycle modeling: learn → act → remember → decompose mega-goals (e.g., design novel anticancer molecules) into thousands of executable subtasks. |
| 2. Autonomous Agent Society | Multi-agent Simulation & Coordination | Build scalable societies of specialized agents (“digital employees”) that debate, review code, allocate resources, and self-optimize workflows. |
| 3. Fully Self-Training | Synthetic Data & Self-Play | Deploy AI-driven data factories and adversarial self-play loops inside secure sandboxes — enabling autonomous model evolution beyond human data limits. |
| 4. Extreme Safety Governance | Mechanistic Interpretability & Value Alignment | Embed ethics, law, and national strategy directly into model value functions; invest ¥10B+ in mechanistic interpretability to transform black-box systems into transparent, auditable, and governable AI. |
🌐 Openness as a Foundational Principle
Parallel to “touching high,” Zhipu doubles down on open access and democratization:
- GLM-5.2 is released under the MIT License, supporting 1M context windows, fully open for download, deployment, and commercial use — no entity restrictions.
- Tang Jie affirms: “Frontier intelligence belongs not to the few, but to all developers — and safety thrives not behind walls, but in sunlight, collaboration, and shared stewardship.”
This dual-track strategy — upward ambition + downward accessibility — defines Zhipu’s unique position: pushing boundaries while paving roads for global builders.
📜 Full Letter Excerpt: Closing Vision
“Why push upward now — when others accelerate monetization? Because true pioneers don’t climb mountains to plant flags — they build roads so everyone can ascend.
We are building that road — high enough to safeguard our nation, wide enough to welcome every developer, and bold enough to reach for cosmic understanding.
The great wave has arrived. And Zhipu will meet it — reaching higher, together.
— Tang Jie, Founder of Zhipu AI
July 11, 2026
Source: Smart Emergence WeChat Official Account