The STAR Market’s fifth set of standards clears the way for large models, accelerating Zhiyuan and MiniMax’s return to the A-share market.

Yesterday (June 17), the Shanghai Stock Exchange issued “Guideline No. 10 for the Application of Issuance and Listing Review Rules,” summed up in one core sentence: Artificial intelligence large-model enterprises can now use the fifth set of listing standards for the STAR Market.

This means that base-model companies that haven’t yet achieved stable revenue—but have burned cash to build strong technological barriers—can finally knock on the A‑share market’s door legitimately. The most direct beneficiaries are Zhipu and MiniMax, which are already listed in Hong Kong and are waiting in line to “return to A.” On the same day, Zhipu’s STAR Market IPO guidance status changed to “guidance acceptance,” perfectly in sync with the pace.

---

What exactly does this new rule open up?

The fifth set of STAR Market standards has been familiar to the market over the past few years—mainly describing biotech and integrated circuit companies: huge R&D investment, long payoff cycles, no short-term stable revenue in sight, but solid technical value. The logic for using the fifth set of standards is simple: profitability is not a hard requirement—technology and market potential are.

Over the past year, with the STAR Growth Segment implemented and the “1+6” reform advancing, the fifth set of standards has already expanded its scope to frontier tracks such as commercial space and low-altitude economy. Now it’s the turn of large models—not unexpected, but the timing is crucial.

The guideline draws several lines:

• Main business must be “independent R&D of AI large models, model services, or model applications”—this covers both general base models and industry vertical models;
• Industry position must be prominent, occupying an important place in the value chain;
• Phased achievements, the toughest requirement: At the time of filing, at least one large-model product must have been launched, released, and achieved scaled application;
• Market space must be substantial, with a clear target market, sufficient potential demand, and strong future growth prospects.

The most worth unpacking is “scaled application.” The SSE’s subtext is: You don’t have to be making money, but you must prove your model is truly being used, not just a press‑conference PPT. This filters out pure PR companies while giving players with real product deployments a way around the “hard revenue indicator.” This standard is more pragmatic than the market had expected.

---

Why must large-model companies “return to A”?

On the surface, Zhipu and MiniMax are already listed in Hong Kong, so in theory they have the financing channels they need. Why go through the trouble of listing in A‑shares?

The answer lies in the nature of the large‑model business.

It is increasingly like a form of capital‑intensive strategic infrastructure:

• Pre‑training and fine‑tuning cycles are short, but each round burns massive computing power;
• Computing clusters are expanding constantly—H‑cards, domestic cards, inference clusters all need investment;
• Data governance and compliance systems are ongoing expenses;
• The density of top talent has hit the ceiling—an annual salary of a million yuan is just the entry ticket.

None of these projects are “one‑and‑done.” They require a company to always have large reserves of deployable cash. Under traditional A‑share rules requiring profitability or stable revenue, base‑model companies face an awkward choice: sacrifice technological frontier for short‑term profit, or abandon the domestic capital market. The fifth set’s openness rips up that dilemma.

A more practical layer is capital structure. A‑shares bring together long‑term funds such as social security, industrial funds, and local industrial capital, with higher recognition of quality AI assets. Having dual “A + H” listings means holding both Hong Kong’s international financing channel and A‑shares’ domestic industrial‑capital platform. For companies burning cash long‑term to iterate, these two platforms are not alternatives—they are twin engines.

---

The hand Zhipu and MiniMax are playing

These two leading companies getting “first dibs” aligns with their product timing.

Zhipu launched and open‑sourced GLM‑5.2 overseas in the early hours of June 17. On the global front‑end development evaluation system Code Arena—with a million users participating in blind tests—GLM‑5.2 ranked first among globally available models. This is a ranking based on real user votes, with more weight than pure academic benchmarks. Zhipu intends to raise no more than 15 billion yuan in A‑shares, with clear uses—general base large model, MaaS one‑stop service platform, and supplemental working capital. The 15 billion figure essentially serves as ammunition for the “computing‑power arms race.”

MiniMax is moving faster. On May 31 it announced “listing” on the Hong Kong Stock Exchange, with fundraising amount undisclosed. On the tech front, its recently launched M3 large model uses the self‑developed MSA sparse‑attention architecture, creating differentiated barriers in multimodal, agent, and code‑engineering tracks. In third‑party evaluations, M3 ranked in the global top tier on multiple indicators. The significance of MSA goes beyond scores—sparse attention reduces reasoning costs in long‑context and Agent scenarios, forming MiniMax’s engineering trump card for its Agent route.

By the way, both GLM and MiniMax M series are already integrated into OpenAI Hub; developers can use the same Key to switch between GPT, Claude, Gemini, DeepSeek, GLM, and MiniMax for comparative testing, without swapping SDKs across platforms.

---

More beneficiaries than just these two

After the guideline’s release, the market quickly sketched a potential beneficiary list:

• General large model direction: DeepSeek, Moonshot AI (Kimi), StepVerse
• Industry verticals: Baichuan Intelligence (medical), Mobius AI (edge AI)

One detail worth noting—the guideline explicitly includes “model applications” in the main‑business scope. This means not only base‑model builders can qualify; companies doing industry MaaS, edge AI, or vertical Agents could also have opportunities. Of course, they must meet “prominent industry position” and “scaled application.”

Following this framework, we are likely to see clear differentiation:

1. Top‑tier base‑model vendors push for A + H dual listings, raising funds into the ten‑billion‑yuan range;
2. Second‑tier general‑model companies face a choice: keep burning cash for base models, or fold into industry‑model companies to seek listing eligibility;
3. Vertical application layer could benefit instead—A‑shares naturally accept AI companies with “clear commercialization path” more readily.

---

A deeper shift: from “tech race” to “deployment race”

Hu Yanping of Shanghai University of Finance and Economics has a view I agree with: two leading companies “returning to A” together will accelerate ecosystem competition differentiation; the old approach of closed‑door R&D with heavy tech focus and light revenue will be unsustainable.

Translated into developer‑friendly terms: in the coming period, domestic large‑model companies will be driven by the capital market toward “industry deployment + commercial monetization.” Base‑model iteration won’t stop, but in parallel will be standardization of industry MaaS, vertical‑scenario solutions, and clear monetization paths.

For application developers, this is actually good news:

• API pricing: Primary‑market investors need to promise growth to the secondary market; revenue will come from either higher prices or higher volume—most will choose the latter, so price wars may continue;
• Industry models: Aggressive investment will appear in fields like medical, legal, financial, and education;
• MaaS infrastructure: One‑stop service platforms will become standard—Zhipu’s 15 billion fundraising includes a specific item for this, implying lowered integration thresholds;
• Agent/toolchains: MiniMax M3’s sparse‑attention route will significantly lower reasoning costs in Agent scenarios, reducing engineering difficulty for long‑context applications.

---

Institutional details: the fifth set’s hard constraints

Back to the guideline itself—several engineering details are worth developer attention:

• The definition of “scaled application” is flexible, but based on previous STAR Market review practices, actual paying‑customer numbers, API call volume, active users, and enterprise‑contract values will likely become core review metrics. This means top‑tier model vendors will definitely push for stronger ToB and ToC usage data.
• The guideline’s “obtaining approval from relevant national departments” implicitly refers to generative AI service registration. Models without registration cannot appear in filing materials—this is an implicit constraint on some products still in grey‑release.
• “Significant technical advantage” in reviews usually requires third‑party evaluation endorsements. Rankings like Code Arena, SuperCLUE, C‑Eval will carry more weight.

---

Conclusion: the “final‑round” war chest

From the industry’s pace, the large‑model competition has tightened to the “final round” with a handful of players. Expanding the fifth set of STAR Market standards is essentially giving these top companies a more ample, sustainable war chest.

Overseas markets provide valuations and international channels; domestic markets provide long‑term capital and industrial‑chain synergy. Running both sides in parallel is the only way domestic large models can truly form closed loops in capital and supply chains.

In the short term, the most direct impact is on Zhipu and MiniMax; in the medium term, it’s on the capital structure of the entire domestic large‑model industry; in the long term, it’s on the cost curve and iteration pace behind the API developers use daily.

For developers, this news is not just about “who’s listing.” The more important thing to watch in the next six months is which model vendors will, in their rush to file, make big moves, cut prices, and roll out industry versions. That’s what this new rule will actually look like in your workflow.

---

References

• Shanghai Stock Exchange: AI Large‑Model Enterprises Eligible for Fifth Set of STAR Market Standards – ITHome: Original guideline highlights from SSE, including explicit wording on key metrics such as “scaled application.”