Zhipu's Tang Jie Internal Letter: Launching 'Touch High' Plan, Aiming for Superintelligence in Two Years

On July 11, 2026, Zhipu founder Tang Jie released an internal letter titled 'The Giant Wave Has Arrived,' announcing the launch of a two-year 'Touch High' plan focusing on long-horizon tasks, autonomous agent systems, self-evolution, and extreme safety governance, directly targeting AGI and even ASI.

Background: IPO Reset, Counter-Intuitive Strategy

Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, at an issue price of HKD 116.2. The stock later surged to a high of HKD 2,980, with a market cap exceeding HKD 1.3 trillion. On the listing day, Tang Jie announced a full return to foundational model research, abandoning short-term commercialization gains. Tang emphasized in the letter that Zhipu's core competitiveness stems from a methodology of 'essence, counter-intuition, and focus,' having previously bet on thousand-billion-parameter models (GLM-130B) and coding capabilities, the latter driving GLM-5.2 to match or surpass Claude Opus 4.8 and GPT-5.5 on multiple benchmarks.

Key Details: Four Engines and Three Mountains

The 'Touch High' plan comprises four core engines:

• Long-Horizon Tasks: Develop a new generation of memory architecture enabling models to autonomously decompose and execute complex tasks spanning weeks to years.
• Autonomous Agent Systems: Build an agent society of thousands of 'digital employees' achieving 24/7 autonomous collaboration, aiming to transition from 'one-person companies' to 'no-person companies.'
• Self-Evolution: Models autonomously write code, synthesize data, and train themselves, forming a recursive self-improvement loop.
• Extreme Safety Governance: Invest tens of billions in mechanistic interpretability to make black-box systems transparent.

Tang pointed out that reaching AGI requires overcoming three mountains: memory (long-term context and RAG approaching prototype), continuous learning (increasing model iteration frequency), and self-judgment (frontier models showing early signs). These challenges are gradually being resolved.

Reactions and Data

• Market Performance: Zhipu's 2025 financial report showed MaaS platform ARR reaching RMB 1.7 billion, a 60-fold increase in one year. In June 2026, Zhipu's market cap entered the 'trillion HKD club,' nearly three times that of Baidu. After the stock lock-up expired on July 8, the share price rose instead of falling, and a follow-on placement raised approximately HKD 31.4 billion the next day, setting a record for a single AI company placement in Hong Kong this year.
• Technical Benchmarking: GLM-5.2 is fully open-sourced under the MIT license, supporting 1M context, and is widely recognized as reaching the capability boundary of overseas frontier models.
• Industry Views: Google DeepMind's report 'From AGI to ASI' noted that even if individual model capabilities stagnate, if AGI instances grow tenfold annually, one hundred million agents at the group level would equate to ASI in five years.

Impact and Outlook

Tang Jie predicted that AI will gradually learn self-awareness, emotion, and even consciousness, moving from perception to cognition to general intelligence, and ultimately to ASI. This trend will reshape application layers, operating systems (LLM OS), and even the von Neumann architecture, affecting industries like finance, law, and e-commerce. Zhipu has chosen a dual path of 'reaching high upward and benefiting downward,' advancing technology while opening the ecosystem and embedding ethical norms into the model's foundation. Tang concluded the letter: 'Not reaching the summit is failure.'