Who are the Chinese labs behind today's open models, and how do they differ?

Much of the open-model race now runs through China. In a July CBC report, the top five models on OpenRouter's usage leaderboard were all open-weight models from Chinese companies. In September, Artificial Analysis's index put Xiaomi's MiMo-V2.6-Pro at 46, GLM-5.3 at 45 and Kimi K3 at 44, the three highest open-weight scores on that board at that date, according to Unite.AI.

Open weight means the trained numbers of a model can be downloaded. These companies publish them for many reasons: to attract developers, to sell hosted versions, to build a reputation abroad and to win customers who won't send data to a foreign service. Their licences show how those aims are changing.

The six labs at a glance

A parameter is one learned number, and active parameters are the ones that work on each token, a small piece of text, in a mixture-of-experts design. Total parameters set the memory you need. Other Chinese groups, including StepFun and Tencent, also release models, but these six lead most recent comparisons. Rankings shift quickly.

What sets each one apart

DeepSeek concentrates on efficiency and writes it up in detail. Its V4.1-Flash paper is about shrinking the memory used by long conversations. It also keeps the simplest terms of the group: MIT, for its flagships.

Alibaba's Qwen has the broadest range, from small open models people run at home to very large ones. Its Qwen3.8-27B is under Apache 2.0, with an 18 GB Ollama download, while the largest carries a custom licence.

Moonshot pushes size. K3's 2.8 trillion parameters are the largest here. Earlier K2 models used a light "Modified MIT", and K3 added conditions for hosting businesses.

Z.ai has often built routes onto Huawei's Ascend chips and other Chinese hardware, and it has been on the US Entity List since January 2025. GLM-5.3 moved from MIT to a custom licence, while its smaller GLM-5.3-Flash stayed MIT.

MiniMax is focused on long context, with its Sparse Attention design, and on multimedia. Its Community License asks the most of small users: attribution and a notice for commercial use, plus a ban on military uses.

Xiaomi came to the field later. Its MiMo-V2.6 series is under MIT and accepts text, images, video and audio. It also published a 9 billion parameter model, distilled and based on Qwen, a good example of how these labs build on each other's work.

How the licences compare

Licences are where these labs differ most, and terms have moved toward more conditions at the top end. Here's what the actual files say:

Alibaba's clause deserves a note, because "AI Work Assistant" extends to independent products mainly for AI-assisted coding or office productivity. Build the next coding assistant on Qwen3.8-Max and, if your group passes the threshold, that clause may apply to you, even though you sell no API. So if you plan a commercial product, read the file before choosing a model.

Is it safe to use a Chinese open model?

It depends on what you mean by safe, so separate the questions.

Data. If you download the weights and run them on your own machines or with a host you trust, your prompts do not reach the original lab. If you use that lab's own app or API, its terms and its home country's laws apply. Our guide to trying DeepSeek-V4 covers what that looks like for one service. Canadian bodies with data-residency rules should treat a lab's app and its downloadable weights as entirely different cases.

Content and behaviour. A model reflects its training and the rules around its maker, so test it on topics that matter to you, including politically sensitive ones. Don't assume that an open file has been audited. Any model given access to tools can do damage if it is misled, so limit its permissions and review what it does.

Law and procurement. Export rules, trade lists and purchasing policies may apply to your organization. Some US officials and companies accuse certain labs of distilling American models, which those labs deny. Those disputes remain unsettled in public. A regulated Canadian organization should ask its legal team and check its own government's guidance.

None of these questions has one answer for everybody. A hobbyist running a small model at home faces little risk from the lab. A hospital or a ministry needs a documented review, whichever country a model comes from.