Choosing Among China's LLMs: DeepSeek, Qwen, Kimi, GLM and Four More, From a Hong Kong School's Side of the Table

Tuesday afternoon, four people round the table in the vice-principal's office, and the IT coordinator slides over a printout of the DeepSeek pricing page. "Fifteen cents per million input tokens against GPT-6 Astra's ten dollars. Put that in the estimates and the finance committee approves it in a minute." The vice-principal reads it twice and asks the only question that matters: "A Form 4 composition uploaded to this — which server is it sitting on tonight?"

China's eight major providers offer the industry's most aggressive pricing and its most permissive weight licences in 2026, and that is a real opportunity for schools. But nobody signs off on a price per million tokens; they sign off on where a student's writing lands after it leaves the campus network, and whose law governs it there.

Where the low prices come from

Start here, or every price comparison downstream is guesswork.

Almost all of the 2026 Chinese flagships are mixture-of-experts designs: the total parameter count is enormous, but only a slice wakes up per token. DeepSeek-V4.1-Flash carries 552B parameters and activates 8B at prefill and 16B at decode. MiniMax M3 carries about 428B and activates roughly 23B per token. GLM-5.3-Flash is 320B total against 18B active; Hunyuan Hy4 Preview is 770B total against 49B active. The vendor's electricity bill tracks the active slice, which is how the rate cards land at a fraction of the American flagships.

Two things then push a school's actual invoice back up.

Time of day. DeepSeek's $0.15 / $0.60 is the off-peak rate and doubles at peak. School-day daytime usage in Hong Kong overlaps the mainland peak window, so cost estimates should assume the doubled tier.

Promotional windows. 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 against a $0.15 / $0.50 list. A three-year budget built on promotional rates assumes no vendor ever raises prices. Kimi K3 sits at $3 / $15, but a cache hit costs $0.30, which changes the arithmetic if the same material is read repeatedly.

The prettiest number on a pricing page is almost never the number the school pays in year three.

Weights versus endpoints

This is the distinction that gets lost fastest, usually under the question "is DeepSeek any good".

Weights are an asset you can carry away; an endpoint is a service you keep renewing. Put the weights in the school server room and nothing leaves the campus network, with hardware and maintenance on your side. Call the official API and hardware stops being your problem, at the cost of handing over the content on every call. Same model name, entirely different privacy assessment.

Licences differ just as materially. DeepSeek ships under MIT and Tencent's Hunyuan Hy4 Preview under Apache 2.0 — the two most permissive, and effectively worry-free for a school's own use. The open build of Qwen's flagship uses a bespoke licence, omits image input, and requires a separate commercial licence above US$50M in annual revenue; a single school is unaffected, but a sponsoring body planning to offer its self-hosted system to other organisations needs to read the clauses first. Kimi K3 and MiniMax M3 each carry their own terms. GLM's flagship weights were still unpublished when this article was updated, though GLM-5.2 and GLM-5.3-Flash are MIT. Doubao is fully closed and served only through Volcano Engine; ERNIE's 5.x weights are unpublished, while the 4.5 family is open.

Three buckets

Sorting the eight by what a school can do with them beats sorting them by capability. Context windows first: DeepSeek-V4.1-Flash, Qwen3.8-Max, Kimi K3, GLM-5.3, MiniMax M3 and Hunyuan Hy4 Preview all reach 1,000,000 tokens; Doubao Seed 2.1 Pro is 256,000; ERNIE 5.1 is 128,000.

What fits in a school server room

Only the small end. Qwen3.8-27B is the clearest case: 27.8B parameters, documented to run within 24GB of VRAM, which is one consumer GPU. A step down, qwen2.5:14b runs on hardware a school already owns. The point is not that this bucket is the smartest, but that it turns "student data does not leave the campus network" from a policy sentence into a technical fact.

What only exists as a remote service

Kimi K3 at 2.8T parameters, Hunyuan Hy4 Preview at 770B and GLM-5.2 at 744B all publish weights that are far beyond what a school can host. In practice they occupy the same position as the closed options — Doubao Seed 2.1 Pro (23 June 2026, ¥6 / ¥30, cache hit ¥1.2) and ERNIE 5.1 (August 2026, ¥0.004 / ¥0.018 per 1K tokens within 32K): remote services requiring a cross-border assessment. On the cloud flagship side, Qwen3.8-Max (3 September 2026) is $2 / $6, GLM-5.3 (14 August 2026) is $1.4 / $4.4, and Hunyuan Hy4 Preview (28 August 2026) has published $0.83 input with output not yet announced.

The vocabulary gap

Rarely raised in procurement, reliably raised by a subject teacher two weeks into term. This group is stronger in Simplified Chinese and mainland context, while a Hong Kong school needs Traditional Chinese, Hong Kong usage and local curriculum framing — a different thing. What appears in output is mainland vocabulary, or a mainland curriculum concept used to explain something inside the DSE syllabus. It is the training distribution showing through, and it is handled at the prompt layer: require Traditional Chinese and Hong Kong usage explicitly, and keep a teacher review step.

One caution on rankings: there is no independent, citable measurement of which model's Chinese is best. Vendor-run blind evaluations — Tencent placing Hunyuan Hy4 Preview slightly ahead of GLM-5.3 and Kimi K3 in its own testing — are self-assessments, not evidence for a school evaluation report.

Published weights only mean you are allowed to carry the model away; whether you have anywhere to put it is a separate invoice.

Crossing the border

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 things: whether the purpose stated at collection covers handing data to an overseas AI provider (DPP1); the bar on using personal data for a new purpose without consent (DPP3); and the fact that the school remains the data user and keeps the security obligation, because outsourcing processing does not transfer it (DPP4).

Worth saying plainly: this is not a China-only problem. A US vendor is also a cross-border transfer; what differs is the school's assessment of the two legal frameworks and the contractual protection on offer. For deployment location and ownership, see on-premises versus cloud deployment.

Three lines in the terms

Most mainland AI services frame their user agreements under mainland regulation. An IT coordinator should be able to point at three specific things in the contract: how long data is retained, whether submitted content is used to train the model, and the mechanism and timeframe for deletion on request. If it is not written down, treat it as absent.

Splitting the workload

Every comparison above resolves into one action: sort the school's AI tasks into two piles, then decide the route for each.

Anything that touches student personal data stays on site. Composition feedback, SEN case summaries, marks analysis, parent letters. Edor.ai's on-premises deployment is feature-identical to cloud and supports LoRA / QLoRA fine-tuning and INT4 / INT8 quantisation on a single consumer GPU; qwen2.5:14b on the suggested list exists for exactly this pile. For hardware, cost and payback, see open weights versus closed APIs.

Anything that does not goes to an assessed cloud provider. Rewriting resources, question variants, public data, grammar explanations — these run on the natively supported OpenAI, Azure OpenAI or Anthropic, where speed and uptime beat a school server.

Individual teacher trials run on de-identified material only. Testing Kimi's long-document handling on a personal account is fine, provided there are no student names, class numbers, marks, SEN flags or parent details. Put that boundary in the policy rather than leaving it to judgement in the moment.

Common misreadings

• The claim: open weights mean the school can use the model however it likes. What actually happens: licences vary sharply. MIT (DeepSeek) and Apache 2.0 (Hunyuan Hy4 Preview) are close to unrestricted, but the open Qwen flagship build, Kimi K3 and MiniMax M3 use bespoke terms with commercial thresholds that have to be read clause by clause.
• The claim: self-hosting an open model settles the AI safety question. What actually happens: it settles data residency only. A self-hosted model still states wrong things confidently, and without the cloud vendor's content-safety classifier, gating, monitoring and the audit trail become the school's job.
• The claim: a Chinese model is naturally the right fit for a Hong Kong school. What actually happens: the strength is in Simplified Chinese and mainland context. Traditional Chinese, Hong Kong usage and local curriculum framing come from the system prompt and teacher review, not from switching vendors.

Three things to try on Monday

1. Take the list of AI tools already in use at your school and write one word beside each: weights or endpoint. The ones you cannot classify are the ones to ask the vendor about first.
2. Give one de-identified piece of teaching material to a cloud model and a local model, same task, and print both outputs for the next panel meeting. The question is not which won, but whether the gap justifies a hardware line in the budget.
3. Add one sentence to the AI use policy short enough to pin on the staff room door: material containing student names, class numbers, marks or SEN flags is processed on the school system only.