China’s open AI stack expands from model weights to agent infrastructure
New downloadable models, a national supercomputing deployment and an open agent framework show Chinese developers competing across the full stack, from weights and licenses to execution infrastructure.
Artificial Intelligence··Morning
Downloadable weights under open licences
Alibaba’s Qwen team released the Qwen 3.8 family under the Apache 2.0 licence, with weights available on Hugging Face and ModelScope. The family includes a multimodal dense model with 27 billion parameters that handles text, images, video, diagrams and documents. Qwen says this model surpasses the larger Qwen3.7-Plus on coding and office tasks, but the comparison comes from the team and the release names no independent evaluation. Its thinking mode can be changed for each query. Zhipu AI has meanwhile released GLM-5.3, built on the same base as GLM-5.2 with gains attributed solely to longer post-training. Zhipu calls it the strongest open-weights coding model and says it expects the weights to follow within two weeks after security reviews. Until then, it is available through the GLM Coding Plan and works with ZCode, Claude Code and OpenCode. Zhipu also says training on vulnerability-detection data helped it find 2,436 flaws across 269 projects with Chinese security teams. Those figures are company measurements without a published independent check. The two releases therefore offer different forms of openness today: Qwen supplies downloadable weights under a stated licence, while Zhipu supplies subscription access and promises the weights later.[2], [3]
National supercomputing as a distribution rail
China’s National Supercomputing Internet has placed the official DeepSeek-V4-Pro-0813 release and the DeepSeek Harness agent framework on its platform. Users can download the model and source code for private deployment, distributed inference and agent development. The platform says its pool combines supercomputing and AI computing across 100,000 accelerators, supports the model life cycle from training through deployment and hosts more than 1,700 open-source models in its AI community. These capacity figures are the platform’s account. Harness entered developer preview on August 13 under the MIT licence. Its plugin-based design lets developers replace model, tool, skill and session components and offers four operating modes. The platform also compares V4 Pro with foreign closed-source models, though that performance comparison remains its own assertion. Wang Peng of the Beijing Academy of Social Sciences says connecting the model to national computing infrastructure strengthens compatibility with domestic systems. The placement gives developers a state-linked route from downloading code and weights to running distributed inference, extending distribution beyond commercial model hubs.[1]
From model intelligence to software control
DeepSeek’s Harness preview turns language models into agents that can operate external software, run code and complete complex tasks. Developers can replace its modular components and intervene directly in the agent’s decisions. The four modes divide general work, commands spanning several applications, experimental custom tools and isolated testing. DeepSeek hired Cui Tianyu, formerly of Jane Street, in March 2026 to lead the new Harness unit. The South China Morning Post describes tools of this kind as an increasingly important AI layer, shifting competition toward the quality of an agent’s connection to real-world software. That shift changes what an open release must provide. Qwen already distributes Apache-licensed weights; Zhipu says its coding weights will follow after review; and the National Supercomputing Internet couples V4 Pro and Harness with domestic computing capacity. Across these releases, Chinese developers are competing over licences, download routes, computing access and agent execution as well as model capability. Their shared timing does not establish a coordinated industrial strategy: the products could still be separate, marketing-driven launches. Yet each addresses a successive obstacle between obtaining model weights and using a model to control software, which makes the stack itself the practical field of competition.[4], [1], [2], [3]
