China’s open-model contest is becoming a platform contest. The question is no longer only which laboratory produces a strong model, but which infrastructure can help developers find, adapt, evaluate and deploy models at scale. As of the September 29 reporting examined here, ModelScope, launched by Alibaba, and Shenzhen-based MoArk represent two different answers to that question.
ModelScope has the larger documented footprint. Alibaba launched it in 2022, and Rest of World reported that the platform had more than 170,000 models. A separate report from Science and Technology Daily described ModelScope users rising from 16 million to 25 million and its model count increasing from 70,000 to 170,000 over nine months. Those figures describe a fast-expanding repository and user base, not a final measure of model quality or market dominance.
MoArk is positioning itself as a competing open-source hub. The company says it launched in 2023 under Shenzhen AOSI Research Intelligent Technology and offers models, data sets, scenarios and localized services. Rest of World reported more than 20,000 MoArk models and described Chinese-chip adaptation work as claims made in an interview with CEO Xu Yong. Those claims matter to the platform contest, but they remain company-attributed rather than independently verified.
ModelScope’s Scale Turns a Repository Into Infrastructure
ModelScope’s advantage is the breadth of the surrounding stack. Its GitHub repository describes model and data hubs alongside inference, training and evaluation capabilities across multiple technical domains. That combination makes the service more than a catalog: it can provide several steps in the workflow that turns an available model into a usable application.
Science and Technology Daily reported that ModelScope supplies models, data sets, toolchains, fine-tuning and secondary-development support. The same report identified more than 34,000 developer-derived Qwen models. The significance is not simply the count. A platform that hosts derivative work can give developers a place to compare approaches, reuse components and build on existing releases. It can also create a feedback loop in which popular model families attract more adaptations, documentation and practical experimentation.
The user numbers add another dimension. Growth from 16 million to 25 million users over nine months, as reported by Science and Technology Daily, suggests that ModelScope is trying to serve a broad development community rather than a narrow research audience. Yet user growth does not show how active those accounts are, nor does it establish that every listed model is maintained or production-ready. The defensible conclusion is narrower: ModelScope has assembled substantial platform scale and a wide set of tools around open models.
ModelScope’s documented structure links discovery with model and data hubs, inference, training, and evaluation. For related archive context, readers can consult EastFrontier’s China open-source AI strategy coverage.
MoArk’s Challenge Rests on Localization and Attribution
MoArk’s proposition is less about matching ModelScope’s reported size and more about offering a localized alternative. According to its About page, the platform provides open-source models, data sets, scenarios and localized services. That combination points toward users who need support adapting open models to specific industries, deployments or hardware environments rather than simply downloading a checkpoint.
The hardware angle is important, but it must be stated carefully. Rest of World attributed MoArk’s Chinese-chip adaptation work to claims from CEO Xu Yong. The available record does not independently verify the more-than-20,000 model figure or establish universal support across Chinese processors. It also does not show that MoArk’s approach is technically superior to ModelScope’s. The strategic claim is therefore about direction: MoArk presents localization as part of its platform identity, while ModelScope’s documented strength is ecosystem scale and integrated development tooling.
That distinction could matter as developers evaluate the total cost of using an open model. A repository can reduce search and experimentation time. Localized services may reduce the work required to adapt a model to a particular scenario. Neither advantage is automatic. Developers still need to assess evaluation results, maintenance, compatibility and the practical quality of support, none of which the fact record establishes for MoArk.
The contest also shows why model counts are an incomplete way to compare platforms. ModelScope’s more than 170,000 models and MoArk’s more-than-20,000 figure, attributed to MoArk, are not equivalent quality measures. A count can include derivatives, specialized versions and projects with different levels of activity. The more useful comparison is how each platform helps developers move from a model listing to a working system.
China’s Open-Model Backbone Will Be Measured in Workflows
The platform that becomes a backbone will likely be judged by repeated developer workflows: finding a suitable model, connecting data, fine-tuning it, evaluating the result and deploying it in a local environment. ModelScope’s repository explicitly covers inference, training and evaluation, while Science and Technology Daily reported its toolchains and secondary-development support. Those are concrete building blocks for that workflow.
MoArk’s stated mix of models, data sets, scenarios and localized services suggests a different route to the same destination. If its chip-adaptation claims translate into reliable deployments, the platform could appeal to organizations that need help aligning open models with domestic hardware. But the current evidence supports only that MoArk is making this positioning claim, not that the deployments have achieved a particular scale or technical result.
For adjacent EastFrontier context, see the DeepSeek-Huawei Ascend software-layer article. The present fact card establishes the documented ModelScope and MoArk platform features, not the broader relationship between every platform and hardware stack.
For Alibaba, ModelScope’s reported scale can help consolidate developer attention around a platform connected to a major technology company. For MoArk, the opportunity is to compete where a smaller platform can differentiate: scenario support, localization and hardware adaptation. The available facts do not identify a winner, and they do not establish that either service has become China’s universal open-model backbone.
The September 29 picture is therefore best understood as an early strategic contest between two platform models. ModelScope brings documented scale, a broad repository and integrated development functions. MoArk claims a growing catalog and a localized service approach, including Chinese-chip adaptation. The next meaningful test will be whether either platform can turn those ingredients into dependable, repeatable workflows for developers across China’s open-model ecosystem.