Chinese open-weight AI

Hi, I’m Mel. Chinese AI no longer means one DeepSeek app. It includes a growing group of models that developers around the world can download and operate themselves. We’ll look at four leading families—Qwen, DeepSeek, GLM and Kimi—and the questions they raise about capability, cost, control and Australian policy.

These are usually called open-weight models. The weights are the trained settings that let a model respond. Releasing them means others can download the model and, subject to its licence, run or adapt it. That is not always the same as fully open source. The training data and complete development process may remain private. Open-weight describes access to the finished model, not everything behind it.

Four families show the range. Alibaba’s Qwen has the broadest international developer ecosystem, many sizes and strong multilingual support. DeepSeek became known for reasoning, efficiency and lower-cost access. Zhipu AI, also known as Z.ai, develops GLM as a balanced general-purpose family. Moonshot AI’s Kimi focuses strongly on coding, tools and long-context work. Their popularity cannot be reduced to one reliable league table, but each has attracted use beyond China.

Qwen is often the practical starting point because its range includes smaller models and it works with familiar local and server tools. DeepSeek offers reasoning models and smaller distilled versions. GLM combines reasoning, coding and agent-style tasks. Kimi’s frontier models are designed for demanding coding and multi-step tool use, but their size creates a much higher hardware barrier. The right family depends on the job and the infrastructure available.

Downloadable does not mean unrestricted or free to operate. Licences vary by release—from permissive Apache or MIT terms to modified or custom conditions. Small, compressed models may run on a capable personal computer. Frontier models can require several expensive accelerators or specialist cloud hosting. Open weights remove one access barrier, while shifting operating cost, security, maintenance and licence compliance to whoever runs the model.

This changed the market. Developers gained capable alternatives they could run, adapt or obtain from competing hosts. DeepSeek’s low prices made that pressure especially visible. Chinese open-weight models helped intensify competition and accelerate cheaper model access. They were not the only cause—better hardware, more efficient designs and wider global competition also reduced costs.

Model origin is not the same as data location. If you use a company’s hosted assistant or API, prompts and files go to that provider under its terms. If an organisation runs open weights in its own approved environment, it can choose where data stays and who has access. That can improve control, but it is not automatic privacy. The deployment, software supply chain, logging and security still need to be managed.

Australia’s government restriction applies specifically to DeepSeek products. It does not cover every Chinese model. PSPF Direction 001-2025 prohibits DeepSeek products, apps and web services on covered Commonwealth systems and devices because of what it calls an unacceptable security risk. It does not bind private citizens or private businesses. It also excludes qualifying locally deployed open models with inspectable code and safeguards. Qwen, GLM and Kimi were not named in that direction, although government users must still assess providers, hosting, foreign influence and information sensitivity.

These models widened access to capable AI, but open does not mean simple. Before choosing one, ask what it can do, what its licence permits, what it costs to operate, where the data goes and what policy applies. Those five questions turn an interesting model into a responsible decision.