Zhejiang Expands AI-for-Science Push Across Research, Industry, and Public Services

Zhejiang has renewed its push to place artificial intelligence inside scientific research, despite a wide gap between policy ambition and verified results. An RSSHub 36Kr newsflash highlighted the province’s support for scientific foundation models and AI-assisted discovery across multiple disciplines.

The announcement sounds like a new policy turn. It is better understood as the latest signal from a program already moving through several layers of provincial planning. Zhejiang released a dedicated AI-for-science action plan in 2025, then reinforced that direction through its 2026 government agenda.

The real story is therefore not another government endorsement of AI. Zhejiang is trying to convert its unusually dense mix of model developers, universities, laboratories, manufacturers, and open-source communities into shared research infrastructure.

That strategy creates a testable conflict. Policy documents promise broad access, open collaboration, and faster scientific discovery. Real progress depends on difficult work involving research data, computing capacity, scientific validation, safety, and incentives for institutions to cooperate.

The RSSHub 36Kr Flash Reflects a Broader Policy Program

The latest headline matters because it connects Zhejiang’s business environment agenda with an existing AI-for-science implementation plan.

The newsflash says Zhejiang’s provincial leadership supports using AI across scientific discovery and technological innovation. It also describes a broader “AI Plus” program covering industry, consumption, culture, public services, and social governance.

RSSHub is the distribution route associated with the item, while 36Kr is the publishing platform. Neither organization created the policy. The underlying actor is Zhejiang’s provincial government, and the relevant program extends beyond a single media update.

That distinction matters for readers arriving through the primary keyword rsshub 36kr. The flash provides a concise event signal, but it does not show the policy’s full history, targets, or unresolved implementation questions.

Zhejiang’s dedicated science action plan was issued in July 2025 by twelve provincial departments. It covers the period from 2025 through 2027.

The plan defines AI for science as the use of artificial intelligence to support scientific research and accelerate discoveries. It focuses on three provincial priorities: artificial intelligence, life and health, and new materials and energy.

By 2027, Zhejiang wants to establish computing, data, and model foundations for AI-assisted research. Those layers are related, but each creates a different implementation problem.

Computing infrastructure supplies the processing capacity needed to train models, run simulations, and analyze scientific datasets. Data infrastructure covers collection, governance, labeling, access, and reuse across organizations.

Model infrastructure refers to scientific foundation models. These systems are trained on domain-specific information and adapted for tasks such as molecular prediction, materials design, or scientific literature analysis.

The province set several measurable targets for 2027. It plans to cultivate at least four scientific foundation models and create at least eight benchmark application scenarios.

The plan also calls for more than twenty representative cases involving intellectual property for scientific data. It aims to support over 1,000 technology companies through the broader program.

These targets make the initiative more concrete than the short newsflash suggests. They also provide a way to separate completed infrastructure from promotional announcements.

Zhejiang’s 2026 work report adds another layer. It supports scientific foundation model development at Zhejiang Lab and an international open-source community around ModelScope.

The report also calls for integrated scenarios in manufacturing, transportation, and public services. This shows that Zhejiang views scientific models as part of a larger regional AI system.

The province is not starting with an empty institutional map. Hangzhou hosts major model development, cloud computing, e-commerce, robotics, and research organizations.

However, geographic concentration does not automatically produce scientific collaboration. Laboratories, companies, hospitals, and universities often store data under incompatible technical, legal, and institutional rules.

The policy’s first challenge is therefore coordination. Zhejiang must turn nearby organizations into a functioning research network without weakening scientific standards or legitimate data protections.

Zhejiang AI Policy Puts Research Infrastructure Under Pressure

Zhejiang’s plan places the greatest pressure on research institutions that control valuable data but lack incentives to make it reusable.

Scientific AI depends on more than general-purpose language models. Researchers need high-quality datasets, specialized tools, reproducible experiments, and subject experts who can identify plausible but incorrect outputs.

A materials model cannot generate useful discoveries from inconsistent measurements with missing experimental conditions. A biomedical model cannot safely combine hospital data without clear privacy, consent, and access controls.

The province’s data targets acknowledge part of this problem. Yet counting data-related intellectual property cases does not reveal whether scientists can actually discover, understand, and reuse the underlying information.

Research data often arrives in formats designed for one laboratory or instrument. Metadata can be incomplete, terminology can differ, and negative experimental results may never enter shared repositories.

These limitations affect model quality. A scientific foundation model can reproduce gaps and biases within its training material, even when its output appears technically sophisticated.

The problem grows when institutions treat data as a strategic asset. Universities want publication credit, companies protect trade secrets, hospitals protect patient information, and laboratories compete for grants.

Zhejiang’s open-source strategy tries to lower some barriers. Open models can reduce dependence on a single vendor and let researchers inspect or modify parts of a system.

Open access does not eliminate costs. Researchers still need computing resources, data engineering, domain evaluation, cybersecurity, and staff who can maintain the models after a pilot ends.

Small companies face the same tension. Access to model weights can reduce software barriers, but running a specialized system may still require expensive hardware and scarce scientific talent.

The province has already linked AI development with open-source communities and support mechanisms. Local programs have included computing vouchers and token subsidies for some developers.

Such measures can help teams begin experiments. They do not guarantee that a prototype will survive procurement, regulatory review, integration, or independent validation.

Pressure will also reach research funders. They must decide whether to support new AI platforms, individual scientific projects, shared datasets, or replication work.

Funding visible models can produce fast announcements. Funding data cleaning, benchmark design, and failed-experiment documentation produces less publicity but often creates more durable research value.

Universities must also adjust incentives. Researchers are usually rewarded for papers, patents, grants, and citations, not for maintaining datasets or evaluating another team’s model.

An effective Zhejiang AI policy must address that mismatch. Otherwise, institutions will announce collaborations while retaining the behaviors that prevent meaningful sharing.

Developers should watch whether the province publishes technical interfaces, licensing rules, benchmark datasets, and evaluation protocols. These details will determine whether outside teams can participate.

Enterprise buyers should look for evidence that models work across institutions. Performance inside one laboratory says little about reliability under different instruments, populations, or operating conditions.

Knowledge workers should care because the same infrastructure problem appears in ordinary organizations. AI becomes more useful when information is structured, traceable, current, and connected to its original context.

The pressure is therefore not limited to scientists. Zhejiang is testing whether regional policy can coordinate the unglamorous information work that useful AI systems require.

Scientific Foundation Models Need More Than Open Access

The central tradeoff is clear: Zhejiang wants broad model access, but scientific reliability demands tighter controls than ordinary chatbot deployment.

Scientific foundation models learn patterns from large collections of research data. They can then support specialized tasks through adaptation, tool use, or connections to experimental systems.

In materials research, a model might rank candidate compounds before laboratory synthesis. In medicine, it might organize literature or identify patterns for further clinical review.

In energy research, models can assist forecasting, simulation, and system optimization. In earth science, they can process observations that would overwhelm manual analysis.

These uses can reduce search costs and help researchers prioritize experiments. They do not turn model output into scientific evidence.

A prediction becomes useful only after appropriate validation. Depending on the field, that may require controlled experiments, external datasets, physical measurements, peer review, or prospective clinical testing.

This creates a direct tension with rapid deployment. Policy programs often reward the number of models, pilots, users, or participating companies.

Science rewards evidence that survives attempts at falsification. A widely deployed model can still be scientifically weak if teams do not test its assumptions and failure modes.

Open-source access can improve scrutiny because more researchers can inspect components and reproduce results. It can also distribute systems before suitable safeguards and documentation exist.

The correct comparison is not open versus closed in isolation. It is transparent, testable infrastructure versus systems whose limitations remain unclear.

Zhejiang has practical reasons to favor openness. The province contains large technology companies, smaller manufacturers, universities, and independent developers with different levels of technical capacity.

A common model and tool layer can reduce duplicated work. Shared components can also help smaller organizations adapt research systems without training everything from the beginning.

However, the province must define what “open” covers. Model weights, training code, dataset documentation, evaluation results, licenses, and safety reports provide different kinds of access.

Publishing weights without training-data documentation limits scientific interpretability. Publishing a benchmark without representative failure cases can create misleading confidence.

There is also a risk of metric optimization. Teams may tune systems to perform well on public evaluations without improving their reliability in unfamiliar research settings.

Independent evaluation becomes essential. Test designers should be organizationally separate from model developers when the results influence funding, procurement, or public claims.

Recent research supports this caution. A laboratory safety benchmark evaluated nineteen advanced language and vision-language models across scientific safety tasks.

No evaluated model exceeded 70 percent accuracy for hazard identification. The study covered 765 multiple-choice questions, 404 realistic scenarios, and 3,128 open-ended tasks.

Those results do not invalidate AI-assisted research. They show why a fluent model should not control consequential laboratory procedures without expert supervision and specialized safeguards.

A separate risk analysis describes vulnerabilities across models, planning, tools, actions, and memory. These risks become more serious when AI systems can affect physical experiments.

Factual errors can lead researchers toward invalid hypotheses. Tool failures can corrupt analysis, while poorly controlled agents can perform actions beyond the intended scope.

Sensitive scientific data creates another exposure. Biomedical records, proprietary chemical processes, and unpublished discoveries can leak through weak access controls or model interactions.

The Zhejiang plan therefore needs a layered governance model. Low-risk literature assistance should not face the same controls as autonomous chemical or biological experimentation.

Human review must remain meaningful. A scientist who merely approves an opaque recommendation without time or evidence for evaluation does not provide effective oversight.

Traceability is equally important. Researchers should be able to identify which data, model version, prompts, tools, and transformations produced a result.

Without that record, teams cannot reproduce findings or investigate failures. Scientific AI then becomes an output generator instead of a reliable research instrument.

The province’s policy language supports scientific discovery, but it should not be read as proof that scientific foundation models are already discovering validated knowledge.

The models are infrastructure candidates. Their value will depend on evaluation standards, documentation quality, integration with experiments, and the willingness to publish negative results.

The Promise of Faster Discovery Meets a Narrower Research Agenda

AI can increase research output while quietly concentrating attention on questions that existing data makes easiest to study.

This is the most important risk missing from a simple rsshub 36kr summary. Faster research does not automatically produce broader or more original science.

AI systems perform best where training data is abundant, digitized, and compatible. That can direct resources toward established fields with measurable outputs.

Problems involving rare conditions, unusual materials, incomplete archives, or expensive physical experiments may receive less attention. Their data provides weaker signals for current models.

Researchers can also converge on similar tools and datasets. When many teams use the same foundation model, its assumptions can shape which hypotheses appear promising.

A 2026 Nature study examined AI-augmented research across a large collection of natural-science papers. It found increased scientific impact alongside a contraction in research focus.

The lesson is not that scientists should avoid AI. It is that productivity metrics can hide a decline in the diversity of questions being explored.

Zhejiang’s targets emphasize models, application scenarios, data cases, and companies served. Those indicators measure activity, but they do not fully measure scientific originality.

The province should also examine whether funded systems produce novel hypotheses, independent replications, useful negative findings, and discoveries across underrepresented fields.

Scientific diversity requires more than diverse model outputs. It requires different institutions, methods, datasets, and intellectual traditions to retain room for disagreement.

This issue connects directly to funding. If grants favor projects with fast AI-generated results, researchers may avoid slower questions with uncertain datasets.

Procurement can create similar pressure. Institutions may standardize around one model because integration is easier, even when alternative approaches would improve scientific resilience.

Open-source systems can reduce vendor concentration. Yet a nominally open ecosystem can still become concentrated around one dominant model, cloud platform, or benchmark.

Zhejiang’s own technology base makes this concern especially relevant. DeepSeek and Qwen give the region recognizable model assets, while ModelScope provides a distribution and collaboration channel.

These strengths can accelerate adoption. They can also make local institutions less willing to test fundamentally different architectures or research methods.

The appropriate response is not artificial fragmentation. The province should support interoperability, independent benchmarks, and grants that compare multiple technical approaches.

Evaluation must also include domain experts who are not model builders. A model can score well technically while producing outputs that scientists find trivial, unsafe, or impossible to validate.

Manufacturing connections offer one route to stronger testing. Zhejiang’s industrial base creates real settings for materials, process, energy, and quality-control research.

For example, the province established an AI and new materials industry alliance in March 2026. Its stated goals include a provincial materials database and at least three design models.

The alliance also plans to support at least twenty research and commercialization projects. Those projects can test whether shared models move beyond demonstrations into repeatable industrial work.

Commercialization is not the same as scientific success. Still, operating environments expose problems involving inconsistent data, equipment variation, latency, maintenance, and human adoption.

The best evidence will combine scientific and operational measures. A useful materials model should improve candidate selection while preserving traceability and performing across different facilities.

Independent replication matters here. Results reported by a model developer or participating company should be treated as claims until other teams reproduce them.

The same standard applies to government-backed pilots. Official support can provide resources and coordination, but it cannot replace scientific verification.

There is also a broader regional competition. Shanghai, Jiangsu, Guangdong, and other Chinese jurisdictions are building their own AI-for-science programs and application platforms.

This competition can create more experimentation. It can also reward headline targets and duplicated infrastructure instead of compatible standards and shared evidence.

Zhejiang’s advantage will not come from announcing another model. It will come from proving that research institutions and companies can reuse validated systems across organizational boundaries.

That outcome remains uncertain. The current documents establish direction, deadlines, and numerical targets, but they provide limited public evidence about cross-institution performance.

Readers should therefore distinguish three stages: policy commitment, operational deployment, and validated scientific contribution. Zhejiang has clearly reached the first and is progressing through the second.

The third stage requires published evidence. Without it, model counts and pilot announcements remain indicators of activity rather than proof of discovery.

What to Watch After the RSSHub 36Kr Headline

Three signals will show whether Zhejiang is building scientific infrastructure or assembling a collection of short-lived AI pilots.

The first signal is public technical evidence from the province’s scientific foundation models. Zhejiang’s 2027 target of at least four models creates a clear deadline.

Watch for detailed model cards, dataset descriptions, licenses, evaluation methods, failure cases, and independent replication. A model announcement without these materials offers little basis for scientific assessment.

Evidence from multiple institutions would strengthen the policy case. Strong performance reported only by the developer would leave the central verification problem unresolved.

The second signal is whether the planned eight benchmark application scenarios survive beyond funded demonstrations. The strongest cases will connect models with real experiments or operating systems.

Look for repeated use, measured research gains, and adoption outside the original project team. A pilot that ends with its grant would weaken claims about shared infrastructure.

Materials research is a useful early test. Zhejiang already has industrial demand, research organizations, model developers, and a new alliance intended to coordinate data and commercialization.

Life sciences will provide a stricter test because privacy, safety, and clinical evidence raise the cost of mistakes. Progress there should be judged more cautiously.

The third signal is whether Zhejiang publishes workable rules for data access and attribution. The action plan’s twenty data-related intellectual property cases offer a starting point.

Useful rules must explain who can access scientific datasets, under what conditions, and with which audit controls. They should also define credit and compensation for data contributors.

If institutions can contribute information without losing recognition or legitimate commercial rights, the data layer can expand. If rules remain vague, valuable datasets will stay isolated.

These signals should be read together. Models cannot improve without accessible data, and application scenarios cannot scale without trusted models.

The rsshub 36kr headline is therefore a useful alert, not the final evidence. It points readers toward a provincial experiment with unusually broad institutional scope.

For developers, the immediate question is whether Zhejiang releases components that outsiders can test and extend. Access should include documentation, not just downloadable files.

For enterprise buyers, the key question is portability. A system that works only inside one subsidized demonstration creates integration risk rather than reusable capability.

For researchers, the decisive issue is scientific value. Faster literature reviews or candidate rankings matter, but they do not replace original hypotheses and validated findings.

For knowledge workers, Zhejiang’s experiment offers a broader lesson. AI performance depends on the quality, provenance, permissions, and structure of the knowledge surrounding a model.

The province has set numerical targets and tied AI to its innovation agenda. It now needs transparent evidence that those targets produce dependable research capacity.

Will Zhejiang publish enough evaluation data for outside researchers to test its claims, or will progress remain visible mainly through model counts and policy updates? That is the question to follow beyond the next RSSHub 36Kr flash.