China's LLMs Compared: DeepSeek, Qwen, Kimi, GLM and Five More — What Can a Hong Kong School Actually Use?

A Chinese panel head asked an excellent question at a demonstration session: "DeepSeek costs one sixty-seventh of GPT-6 Astra per input token and its Chinese is no worse. Why doesn't your platform just connect to it?" He was right on the arithmetic — on September 2026 published pricing, DeepSeek-V4.1-Flash is $0.15 per 1M input tokens against GPT-6 Astra's $10, exactly 66.7 times cheaper. But the answer to his question is not on the capability table. It is in the question of which jurisdiction a student's composition lands in after it leaves the school.

This article puts all eight Chinese vendors on one table, then opens up the decision that actually matters: data residency and licensing.

Plain-language mechanics: three layers that get conflated

When Hong Kong schools discuss Chinese models, three separate things usually get treated as one.

First, whether the weights are public. Weights are the recipe: publish them and anyone can cook the dish at home. DeepSeek ships under MIT and Tencent's Hunyuan Hy4 Preview under Apache 2.0, the two most permissive open licences. But public is not the same as a free lunch — the open build of Qwen's flagship uses a bespoke licence without image input, and Kimi K3 and MiniMax M3 each carry their own terms.

Second, whether you are using the weights or the API. This is the decisive layer. Call the official API and your data leaves the school for the vendor's servers. Download the weights and run them on a school server and nothing leaves the campus network. The same "DeepSeek" carries wildly different privacy risk under those two uses.

Third, total parameters versus active parameters. Parameter count is the brain-cell count, but nearly every Chinese model in 2026 uses a mixture-of-experts design: DeepSeek-V4.1-Flash holds 552B parameters yet activates only 8B at prefill and 16B at decode; MiniMax M3 holds about 428B and activates roughly 23B per token. That is how the prices get this low. It also explains why "many parameters" does not mean "easy to self-host" — you still need enough VRAM to hold all of them.

Head-to-head: eight Chinese providers

Prices in USD per 1M tokens (input / output) unless noted, from each vendor's official documentation as of September 2026.

Several details in the official documents are easy to miss when comparing prices:

• DeepSeek's $0.15 / $0.60 is the off-peak rate and doubles at peak. A Hong Kong school's peak usage — daytime on school days — overlaps the mainland peak window, so budget at the higher rate.
• GLM has a two-layer structure worth understanding. GLM-5.3-Flash (26 August 2026, 320B total and 18B active parameters) is MIT-licensed at a promotional $0.075 / $0.25 against a $0.15 / $0.50 list price, making it the cheapest of this batch. The GLM-5.3 flagship weights were still unpublished when this article was updated.
• Hunyuan has the cleanest licence. Hy4 Preview publishes 770B total and 49B active parameters under fully permissive Apache 2.0 — clearer terms than the bespoke licences on GLM-5.3 or Kimi K3. The catch is that self-hosting it still needs server-class hardware.
• ERNIE's context is well behind. At 128K tokens it is under an eighth of its contemporaries, and its headline strength — live answers grounded in Baidu search — is exactly the part a Hong Kong school needs to assess most carefully.
• Doubao is fully closed and served only through Volcano Engine. Its pricing is in RMB and its terms and data handling sit under mainland regulation.

On the question of which has the best Chinese, we have no independently measured third-party results we can cite. Vendor-run blind evaluations — such as Tencent placing Hunyuan Hy4 Preview slightly ahead of GLM-5.3 and Kimi K3 in its own testing — are vendor self-assessments and schools should not treat them as independent evidence. From our own school deployments, this group is generally stronger in Simplified Chinese and mainland context, which is not what a Hong Kong school needs: Traditional Chinese, Hong Kong usage and local curriculum framing are a separate matter, and these models readily produce mainland vocabulary or apply mainland curriculum concepts.

The decision that actually matters: where the data goes

This section matters more than the table above it.

All eight serve their official APIs from mainland China. The moment a school calls one directly, student compositions, marks, SEN records or parent contact details leave Hong Kong for another jurisdiction. Under the six Data Protection Principles of the Personal Data (Privacy) Ordinance (Cap. 486), that engages at least three points:

1. Purpose and notification (DPP1): does the purpose stated when the school collected student data cover handing it to an overseas AI provider?
2. Use limitation (DPP3): personal data must not be used for a new purpose beyond the original one without consent.
3. Security (DPP4): the school remains the data user and remains responsible for security; outsourcing processing does not transfer that.

There is a contractual layer too. Most mainland AI services frame their user agreements under mainland regulation, so an IT coordinator should check retention periods, training-use clauses and deletion mechanisms line by line. None of this means Chinese models cannot be used. It means the assessment has to be completed first, and written into the school AI policy and the parent notice.

Worth stressing: this is not unique to Chinese models. Using a US vendor is also a cross-border transfer; what differs is the school's risk assessment of the two legal frameworks and the contractual protections on offer. For the full discussion of deployment and ownership, see on-premises versus cloud deployment.

Limits and blind spots, stated plainly

Open weights do not remove hallucination. Self-host DeepSeek or Qwen and the data is safe, but the model will still state wrong things confidently. Worse, a self-hosted model lacks the content-safety classifiers a cloud vendor runs, so the school must supply its own gating, monitoring and audit trail. Privacy risk and content risk are two different lists.

"Downloadable" is not "runnable". Kimi K3 at 2.8T parameters, Hunyuan Hy4 at 770B and GLM-5.2 at 744B are all far beyond what a school server can host. What is genuinely realistic for schools is the small end: Qwen3.8-27B (27.8B parameters, documented to run in 24GB of VRAM), the Qwen2.5 family, and certain distilled or quantised builds.

Licences carry real commercial restrictions. The open build of Qwen's flagship, for instance, requires a separate commercial licence for providers above US$50M in annual revenue. A school's own use is generally unaffected, but a sponsoring body planning to offer its self-hosted system to other organisations must read the terms first.

Promotional pricing ends. MiniMax M3's $0.30 / $1.20 is promotional against a $0.60 / $2.40 list, and GLM-5.3-Flash's $0.075 / $0.25 likewise. Building a three-year budget on promotional rates is not safe.

Traditional versus Simplified and vocabulary differences need handling at the prompt layer. This is not a defect, just the natural result of the training distribution. In practice the system prompt must explicitly require Traditional Chinese and Hong Kong usage, and the teacher interface must include a review step.

Actionable guidance: three workable routes

Route one (recommended): self-host open weights on site for sensitive data. This suits schools with IT capacity and an explicit "data stays on the campus network" policy. Run a Qwen3.8-27B or qwen2.5:14b class model on a school server through Ollama and have the platform route sensitive tasks to the local model. Edor.ai's on-premises deployment supports LoRA / QLoRA fine-tuning and INT4 / INT8 quantisation on a single consumer GPU. For the cost, hardware and payback arithmetic, see open weights versus closed API.

Route two: a cloud primary provider with an explicit boundary. Everyday teaching tasks go to an assessed cloud provider (the platform natively supports OpenAI, Azure OpenAI and Anthropic), while anything touching student personal data stays on site or does without AI. This is what most schools actually do.

Route three: individual teacher trials, on de-identified material only. A teacher who wants to test Kimi's long-document handling or DeepSeek's mathematical reasoning may use a personal account, provided the material is fully de-identified — no student names, class numbers, marks, SEN flags or parent details. Put that line in the school AI use policy rather than leaving it to individual judgement.

A five-point checklist for the IT coordinator:

1. Which weight licence does this model use? MIT and Apache 2.0 are the most permissive; bespoke licences must be read clause by clause.
2. Are we using the weights or the API? If the API, which jurisdiction does the data enter?
3. Do the terms state explicitly that school data will not be used for training? What is the retention period?
4. Will the hardware hold it? 24GB of VRAM is the practical dividing line for school self-hosting.
5. Is the decision reversible? If the model is replaced, are the knowledge base and historical records affected?

Takeaway

China's models offer the industry's most aggressive pricing and its most permissive weight licences in 2026, and that is a genuine opportunity for schools — but the opportunity sits on the self-hosted open-weights side, not on the "call a mainland API directly" side. The right question for a Hong Kong school is not "is this model good", but "may this data leave the campus network, and if not, which model can I bring inside".

For the individual provider pages, see DeepSeek, Alibaba Qwen, Kimi, Zhipu GLM, MiniMax, Doubao, Tencent Hunyuan and Baidu ERNIE. For how they compare against the US providers, see the 2026 all-model roundup.