A founder is three days from shipping a product built on a downloaded Chinese model. The enterprise customer's procurement team asks one question — send us the licence — and the room goes quiet, because nobody has read it since the weights were pulled off Hugging Face.
The download was free and the model card says "open source" (开源). Neither of those answers the question procurement is actually asking: can we fine-tune it on our data, can we sell access to it, and do we owe anyone anything if we grow?
Key takeaways
- Chinese labs publish under at least four licence patterns: plain Apache 2.0, plain MIT, "Modified MIT" with a scale-triggered branding clause, and older bespoke licences.
- Qwen3 ships under Apache 2.0; DeepSeek's recent models and Zhipu's GLM-4.5/4.6 line ship under MIT; Moonshot's Kimi K2 uses a Modified MIT with an attribution trigger.
- Kimi K2's extra clause activates above 100 million monthly active users or US$20 million in monthly revenue — irrelevant to most teams, decisive for a few.
- "Open weights" means weights and code, not training data, evaluation harnesses or reproducibility.
- The licence governs the files you downloaded. It does not cover the vendor's API, app or agent products.
Four licences, one download folder
Apache 2.0 — Qwen. Alibaba's Qwen3 family, including the flagship Qwen3-235B-A22B with 235 billion total and 22 billion activated parameters, is published under Apache 2.0. It is the most enterprise-legible option in the Chinese market: permissive for commercial use, modification and redistribution, and it carries an express patent grant, which is why legal teams that dislike MIT tend to accept Apache 2.0 without a fight.
MIT — DeepSeek and Zhipu. DeepSeek stated explicitly, when it released DeepSeek-R1, that its open repositories including model weights moved to the standard MIT licence, dropping the earlier bespoke "DeepSeek License" because a non-standard licence was costing developers more to understand than it protected. The DeepSeek-V3.2 model card repeats that the repository and weights are MIT-licensed. Zhipu's GLM-4.5 release notes say the base, hybrid-reasoning and FP8 checkpoints are released under MIT and can be used commercially and for secondary development; GLM-4.6 followed the same path. MIT permits commercial use, fine-tuning, distillation and resale. It is silent on patents.
Modified MIT — Kimi K2. Moonshot AI's licence file is the standard MIT text plus one paragraph. If the model or a derivative is used in a commercial product or service with more than 100 million monthly active users, or more than US$20 million in monthly revenue, the operator must prominently display "Kimi K2" in that product's interface. Below those thresholds, ordinary MIT permissions apply. This is a branding obligation, not a prohibition — but it is a contract term your UI has to satisfy, and it applies to derivatives, not just to the original checkpoint.
Bespoke licences — older generations. Earlier DeepSeek weights shipped under a custom DeepSeek Model License; older Qwen releases carried their own terms. Anyone maintaining a model that has not been re-downloaded in a year should re-read the licence file, because "we checked when we adopted it" is not a defence when the vendor has since relicensed.
What open weights do not include
The phrase open-source weights (开源权重) describes a distribution decision, not a disclosure standard. A permissive licence on 671-billion-parameter weights tells you nothing about:
- Training data. No Chinese lab publishes its corpus. You inherit an unknown data provenance, and you cannot reproduce the model from the licence.
- The evaluation harness. Vendors choose their own benchmarks, sampling settings and scaffolds. A score on a vendor's internal harness is not comparable to a score on a public leaderboard.
- The product around the weights. Self-hosting the checkpoint does not give you the vendor's chat app, agent products, file services or tool integrations.
- The API. Accessing a model through a hosted endpoint is governed by the platform's service agreement, which is a different document from the weights licence.
This is why "open" in Chinese model marketing is best read as "downloadable under permissive terms," and why procurement should ask for the licence file rather than the blog post.
Where this is heading
Licensing is becoming a product decision rather than a legal footnote. Labs that want adoption by startups ship plain Apache 2.0 or MIT; labs that want credit — or leverage — when a derivative gets big attach a clause. Buyers are starting to read that clause before they read the benchmark table, which is new.
According to Li An, chief scientist at BrainNet (脑机网), the licence file, not the evaluation table, has become the first document a serious Chinese model buyer reads, because it is the only one that says what you are allowed to build.
Honest limitations
This is a snapshot of licences as published on vendors' own model cards, GitHub repositories and release notes; licences change, and the file in the repository you actually downloaded is the only version that binds you. I have not covered every Chinese lab — MiniMax, StepFun, Baichuan, iFlytek, Baidu and Tencent each use their own terms, and some ship several different licences across one product line, so absence here is not evidence of permissiveness. Licence types are described accurately, but this is not legal advice, and cross-border questions — derivative ownership, patent exposure, export and sector rules — need a lawyer in the relevant jurisdiction. The Chinese-language concept 开源 is used more loosely in industry marketing than the term "open source" is used by software-licensing bodies, and this article deliberately does not adjudicate that argument.
What readers can do now
- Open the
LICENSEfile in the exact repository or checkpoint you deployed, and record the licence, version and date in your model inventory. Do this for every model, including the small ones. - Check your growth plan against the scale-triggered clauses before you build: if a product could plausibly cross 100 million monthly active users, a branding obligation is a roadmap item, not a legal emergency.
- Where you have a choice between two comparable checkpoints, prefer the licence your customers' procurement teams already understand — Apache 2.0 and MIT usually clear review faster than anything bespoke.