Zhipu co-founder Tang Jie published a full-staff letter on July 11 declaring a two-year “Touch High” plan: pour resources into foundation-model research, keep frontier weights open under MIT, and commit tens of billions of RMB to “mechanical interpretability” safety research. The letter landed days after a lockup expiry sent Zhipu’s Hong Kong-listed stock down more than 19% — and Tang’s response was to double down on the least commercial path available.
🔍 THE BOTTOM LINE
While Anthropic and OpenAI are tightening access to their most capable models, citing risk concerns, Zhipu is moving in the opposite direction: open weights, permissive licensing, and an explicit bet that AGI cannot be reached behind closed doors. The strategic contrast is not subtle. Tang framed it as a civilisational question: “When a technology reaches the level of force capable of altering the course of civilization, safety is no longer an ancillary feature; it becomes the prerequisite for the technology’s permitted existence.” The implication is that openness — not lockdown — is the path to legitimate AI deployment.
What the Letter Actually Says
Tang Jie’s letter, published via Geopolitechs and dated July 11, 2026, is structured in five sections. The core message is a strategic pivot away from short-term application revenue toward what Tang calls the “Touch High Plan” — a two-year investment push concentrated on four engines: long-horizon task execution, autonomous agent systems, fully self-training models, and extreme safety governance.
The timing is deliberate. Zhipu listed on the Hong Kong Stock Exchange on January 8, 2026, and a lockup expiry triggered a 19%+ stock decline. Rather than reassure investors with a commercial roadmap, Tang told staff to “reset to zero” and return to foundation-model research. The letter’s opening line: “Others rang the bell; we reset to zero.”
Tang defines AGI not as a single genius’s intelligence but as “the sum of all human wisdom,” and says the model should be capable of “creating original knowledge on the order of the Theory of Relativity.” That is a high bar — and one no current model meets.
The Open-Weights Bet
The product expression of Tang’s stance is GLM-5.2, released under the MIT license with no restrictions based on entity type. The model supports a one-million-token context window and is available for anyone to download, deploy, and commercialise. As CNBC reported in June, GLM-5.2 lands within a percentage point of Anthropic’s Opus 4.8 on a key agentic benchmark at roughly a fifth of the cost.
This is not a fringe move. Zhipu’s GLM-5.2 sits at third place on the Artificial Analysis Intelligence Index, trailing only Anthropic and OpenAI. The company’s agentic coding harness ZCode 3.0 pulled 368 Hacker News upvotes in 15 hours when it launched. And as we noted in our coverage of GLM 5.2’s margin-collapse implications, the switching cost from frontier API providers to open-weight alternatives is approaching zero.
The open-weights stance puts Zhipu in direct philosophical opposition to the Western frontier labs that Tang references obliquely in his letter. When “the most advanced overseas frontier models have deferred full public release due to risk considerations,” Tang writes, that is a signal — not a curiosity. His conclusion: safety and openness are not in tension; they are prerequisites for each other. “Genuine safety is not built on technological closure and barriers, but on broad co-construction, co-sharing, and oversight conducted in the open.”
The Safety Investment
The fourth engine of the Touch High Plan is where the money goes. Tang says Zhipu will commit “resources in the tens of billions” to “mechanical interpretability” — research aimed at converting opaque model decisions into transparent, auditable logic. The goal is to transform black-box systems into explainable ones.
This is a notable commitment from a Chinese AI lab, and it lands at a moment when the global safety conversation is in flux. OpenAI’s head of safety Johannes Heidecke just departed after a reshuffle. Anthropic continues to withhold its most capable model from full public release. The question Tang is implicitly raising: if you close the model to stay safe, but nobody can inspect the model to verify the safety, what have you actually achieved?
Tang’s framing embeds “human ethics, social norms, and national laws and regulations” as “foundational axioms in the model’s value function” — not as post-hoc compliance patches. Whether Zhipu can actually deliver on that promise at scale is an open question, but the commitment itself is a different posture from the bolt-on safety approach that has dominated the industry.
The Three Peaks to AGI
Tang identifies three technical peaks Zhipu must cross on the road to AGI:
- Long Horizon Task — models that can plan and execute over weeks, months, or years, not just answer instant queries. Tang’s example: a model working “tirelessly at the endpoint of a physics laboratory.”
- Autonomous Agent System — clusters of intelligent agents that collaborate, debate, review code, and schedule resources autonomously. Tang references the shift from “One-Person Company” to “Fully Automated Company.”
- Self-Evolving — AI training AI. Models that write code, clean data, and train themselves. Tang acknowledges this “may consume some compute” but saves “the most precious resource: human labor and time.”
He also cites Google DeepMind’s report “From AGI to ASI,” which argues that even if a single model never exceeds human level, superintelligence may be “squeezed out” by the aggregation of massive compute resources — one hundred million AGI instances within five years, thinking one hundred times more efficiently and replicating experience at zero cost.
The China Context
Tang’s letter is also a geopolitical document. The open-weights stance is a direct challenge to US export controls and the Western strategy of controlling AI capability through chip access. As we’ve covered in our reporting on DeepSeek building its own chips, Chinese labs are finding paths around hardware restrictions. Zhipu’s bet is that the answer to being cut off from chips is not to hoard capability but to open it — to make the weights so widely available that control becomes structurally impossible.
Inside AI News reported in June that Tang told Elon Musk on X that China would match Anthropic’s top model “this year, not in 2027.” Musk had estimated Q1 2027. Tang’s reply: “won’t take that long.” GLM-5.2’s benchmark performance — second globally on Code Arena’s front-end coding benchmark — suggests he may be right.
What This Means for the Frontier
The competitive implications are significant. If Zhipu delivers on the Touch High Plan and continues shipping open-weight models at or near frontier quality, the pricing pressure on closed labs intensifies. GPT-5.6 Sol costs $5/$30 per million tokens. Gemini 3.5 Pro is rumoured to launch July 17 at $1.25/$10. GLM-5.2 is free to download. The margin compression we flagged in June is not a projection — it is happening now.
The deeper question Tang’s letter raises is about the relationship between safety and openness. The Western frontier labs argue that closing access is the responsible move. Tang argues that closing access is the risk move — that genuine safety requires broad scrutiny, not a handful of gatekeepers. It is the most coherent articulation of the open-weights-as-safety position from a major lab leader to date.
Whether Zhipu can execute on a two-year AGI plan while its stock is down 19% and its compute base is constrained by export controls is the practical question. The philosophical question — should frontier AI be open or closed — just got a lot harder to dismiss.
❓ FAQ
Why did Tang Jie publish this letter now? The letter followed a lockup expiry that sent Zhipu’s stock down more than 19%. Rather than reassure investors with a commercial strategy, Tang doubled down on foundation-model research and open weights. The letter is a statement of strategic direction, not a market reaction.
Is GLM-5.2 actually competitive with Western frontier models? On the Artificial Analysis Intelligence Index, GLM-5.2 ranks third globally, behind Anthropic and OpenAI. On Code Arena’s front-end coding benchmark, it ranks second behind Anthropic’s Fable 5. CNBC reported it matches Opus 4.8 on agentic benchmarks at roughly a fifth of the cost. The gap is narrowing, though benchmark performance does not always translate to real-world capability.
What does “mechanical interpretability” mean? It is research aimed at understanding why a model produces a particular output — tracing the internal logic of neural networks so decisions can be audited and explained. Tang frames it as the path from black-box systems to transparent ones, and says Zhipu will commit “resources in the tens of billions” to it.
How does this affect the open-weights debate? Tang’s letter is the most explicit argument from a major lab leader that open weights are a safety feature, not a safety risk. The position directly contradicts the Western frontier lab stance that closing access is the responsible move. It raises the question: if safety requires scrutiny, can closed models ever be verified as safe?
What is the “Touch High” plan? A two-year strategic investment plan focused on four areas: long-horizon task execution, autonomous agent systems, fully self-training models, and safety governance. The plan explicitly deprioritises short-term application revenue in favour of pushing toward AGI.
🔍 THE BOTTOM LINE
Tang Jie’s letter is the clearest articulation yet of the open-weights-as-safety thesis from a major AI lab. The bet is large, the timing is adversarial, and the philosophical stakes are real: if Zhipu can ship frontier-quality open weights while investing tens of billions in interpretability research, the closed-lab argument gets harder to sustain. The AI race just gained a new axis of competition — not just who has the best model, but who has the most defensible theory of how AI should be governed.
