Two Beijing research institutions have used their second-anniversary event to present eight artificial-intelligence projects spanning industrial software, genomics, materials, molecular design, low-precision inference, physical intelligence, and single-cell biology. The September 13 release from Zhongguancun Academy and the Zhongguancun Institute of Artificial Intelligence offers a compact view of how China’s public-interest research groups are trying to connect model work with sectors where scientific and industrial constraints matter as much as benchmark performance.
The two institutions, often described together in Chinese reporting as the Zhongguancun Two Institutes, presented the projects at the Xᴬᴵ Technology Festival. The list includes Modelware, a software-intelligence effort; Carbon, a genomics model and data initiative; FlexRibbon, a drug-design engine; MatOS, a materials-research operating system; and four projects aimed at inference, physical intelligence, cell biology, and generative modeling.
The announcement should not be read as eight finished products arriving at once. Some of the projects have earlier research histories, some have stated future open-source dates, and many of their performance figures come from the institutions themselves. What is new is the decision to present them as a coordinated research portfolio, joining AI for science with the harder problem of domestic software and industrial capability.
Software intelligence is the economic centerpiece
Modelware is the most explicitly strategic item in the group. Beijing News describes it as a super software-intelligence project intended to help with the migration, development, and adaptation problems that have slowed China’s replacement of foreign basic and industrial software. That matters because software autonomy is not achieved by producing a single substitute application. It requires compatibility with a large installed base of code, tools, and workflows.
The institutes say Modelware’s initial results have been open-sourced. That is a concrete step, but the broader claims still need to be judged by outside developers and industrial users. Engineering software has long development cycles and severe reliability requirements. A model that can draft code or interpret a specification may be helpful, yet it must still work within testing, version control, security review, and existing engineering practices.
The project sits beside a broader national effort to move AI beyond consumer chatbots. EastFrontier recently examined China Telecom’s forecast of agent-driven AI infrastructure, which also highlighted the need to connect models to usable systems. Zhongguancun’s contribution is more research-driven, but the policy logic is familiar: use models to shorten the route from a complex problem to usable software.
A portfolio built around science and physical systems
Several of the new releases focus on scientific discovery. Carbon is a genomics large model paired with a planned Carbon Database. FlexRibbon targets the difficult task of designing around flexible molecular interfaces, while MatOS is presented as an autonomous-agent operating system for materials research. These projects address domains where a model must work with specialized data and established scientific constraints rather than general internet text.
The MatOS description is particularly revealing. The institutes say it links industrial needs with research workflows and integrates thousands of datasets, hundreds of calculation methods, and dozens of research software tools. Those numbers are institute-reported, as are examples of efficiency improvements in particular applications. They are useful indicators of ambition, not a universal benchmark for AI-led materials science.
The remaining projects extend the portfolio beyond familiar large-language-model categories. Small Value Domain Intelligence is described as a lossless FP4 low-precision inference framework. PhysBrain uses multimodal human-behavior data to train a physical-intelligence model. HoloCell is presented as a single-cell tri-omics foundation model, and the quotient-space diffusion framework applies symmetry to generative modeling.
Together, they show a research agenda that treats efficiency and physical data as first-class problems. That matters for China’s AI sector, where compute supply remains valuable and scientists want models that can work with measurements, simulation, and laboratory procedures. The Zhengzhou scientific-AI computing cluster illustrates a related effort to build infrastructure for models, agents, and experimental work.
Open-source promises create the next accountability test
The institutions have given some projects a public timeline. Beijing News reported that the Carbon Database is scheduled for open source on September 15, while MatOS is expected to be released around mid-October. Those are plans, not completed releases, and the final scope will matter. An open model or data resource is most useful when researchers can understand its provenance, reproduce its results, and determine how it behaves outside a curated demonstration.
That challenge is especially important in genomics and materials research. These fields depend on data quality, experimental conditions, and domain-specific interpretation. Making a model available is only the first stage. External users will want documentation, evaluation protocols, and clear limits on what the system can claim.
The Zhongguancun release is nonetheless notable because it moves beyond a single headline model. It presents a stack of research directions that includes software, chips, science, biology, and physical intelligence. The common thread is not that AI can replace a scientist or an engineer. It is that models may become another layer of infrastructure for work that already depends on specialized tools and human validation.
For a research ecosystem only two years old, that is an ambitious agenda. The more meaningful question will emerge after the festival: which of these projects attracts outside users, which survives contact with real industrial data, and which can demonstrate value without relying on a carefully controlled showcase. Beijing News’ report provides the primary account of the eight named results and their announced release plans.