Alibaba Qwen (Qwen3.8): the open family most respected for Chinese, and the two routes open to Hong Kong schools

Qwen appears on almost every Chinese-subject procurement slide — yet meetings often read "Qwen is strong" as "we must use the Max cloud API". For a Hong Kong school the real fork is Qwen3.8-27B on a 24GB card inside the campus network, or Max in the cloud as a cross-border transfer that needs a privacy assessment first.

Qwen is Alibaba's model family, shipping each generation as both a cloud-only flagship and downloadable open weights. For Hong Kong primary and secondary schools that structure maps to two costs: cloud for the strongest Chinese and a 1M-token window, self-hosting for simpler compliance and upfront hardware — neither is universally better, only different costs the school prefers to carry.

What the company is actually building

Qwen's rhythm is flagship in the cloud, sizes downward for download. Qwen3.8-Max reached general availability on 3 August 2026 with a 0902 snapshot on 3 September — a 2.4T-parameter MoE activating about 95B per token, text, image and video input, 1M-token context — but the Max endpoint is closed and its weights are not published.

On the open side, Qwen3.8-27B (14 August 2026) is a 27.8B dense vision-language model under Apache 2.0, documented to run in 24GB of VRAM — the size that turns "a Chinese model in school" from theory into a purchase order. Qwen3.8-Flash (26 August 2026) is a 125B total / 6B active small MoE for high-volume batch work. The Qwen2.5 series (19 September 2024) remains the most-used Chinese open generation on Ollama; qwen2.5:14b in our suggestion list belongs to that generation.

One point worth repeating: Qwen's Chinese is strong but training context leans mainland. Output may say "Grade 5" rather than "Primary 5", or "lesson plan" rather than "worksheet" — a context problem fixed with prompt templates, the school knowledge base and teacher review, not automatically by buying a more expensive tier.

Model line-up at a glance

Providers above US$50M annual revenue need a separate commercial licence for 2.4T-A95B; ordinary Hong Kong school use is unlikely to trigger it — not being able to run the weights is the main reason to choose 27B.

Where the money actually goes

Cloud Max / Flash: flagship at $2 / $6, Flash at $0.15 / $0.47 — teachers can run capability tests on material with no personal data at almost no cost. Sending a student's composition or name to the Max API is a cross-border personal-data transfer: either complete a privacy impact assessment with informed parental consent, or self-host instead.

On-premises 27B or 14b: a 24GB card is a purchasable spec, and Apache 2.0 lets a sponsoring body deploy one weight set across several schools with no user cap. Hidden costs are power, cooling, backups and version upgrades — without IT headroom, total cost may not beat a cloud subscription.

Platform stack: even with Qwen on chat, embedding and Cantonese oral practice still need OpenAI or Azure OpenAI. Self-hosting settles Chinese conversation residency, not "zero cross-border for the whole Edor.ai stack".

A 24GB card buys compliance simplification, not automatic access to 2.4T MoE capability — procurement papers should state which size you run, not only "we use Qwen".

What role it plays inside Edor.ai

Qwen is not a native provider. It arrives through local Ollama: run Qwen3.8-27B or qwen2.5:14b on a school server and set the provider to local Ollama in the admin panel. The suggestion list already includes qwen2.5:14b, so an IT coordinator is not starting from zero.

Self-hosted mode

Conversations reach no external cloud. Four-layer gating, teacher monitoring, ai_audit_logs and knowledge-base retrieval work as normal; data belongs to the school and exports in full at any time. Chinese lesson plans, classical–modern comparisons, worksheet variants and long Chinese summaries are usually enough for teachers on 27B without reaching for Max.

Student use

Suited to Chinese reading-comprehension hints and vocabulary explanations in tutor mode only — the platform holds the model to hints rather than answers and keeps conversations a teacher can review. Value-judgement topics in humanities should be teacher-led; answers on sensitive or political topics may be guarded or skewed and must never be the only source.

Insisting on the Qwen3.8-Max cloud API means accepting a cross-border transfer and completing a privacy assessment first. Schools that want output closer to school conventions can fine-tune open weights with LoRA, QLoRA and INT4 on a consumer GPU — the same concepts as the LLM Classroom.

Where it fits and where it does not

Fits: Chinese, General Studies and Chinese History planning; schools with a 24GB card willing to operate Ollama; sponsoring bodies wanting Apache 2.0 replication; student scaffolding in tutor mode with teacher review.

Does not fit: student work on Max cloud before assessment; treating 2.4T open weights as runnable on campus; circulars to parents without terminology calibration; history and current-affairs questions handed entirely to the model; self-hosting assumed "buy once" with no IT staffing.

After school on a Friday, a Chinese panel head tells the IT coordinator: "27B sped up worksheet drafts a lot, but I still change 'trash can' to 'rubbish bin' — I added a Hong Kong usage line to the template." That is a normal rollout rhythm.

Copy these prompts to test:

You are a Hong Kong primary Chinese-language teacher. Design a 30-minute reading and writing activity on the origins and customs of the Mid-Autumn Festival for Primary 5.
Output: one passage under 350 words, a vocabulary table of six terms with example sentences, four comprehension questions, two writing prompts, and three teacher observation points.
Constraints: write everything in Traditional Chinese using Hong Kong conventions, using Hong Kong grade and material names rather than mainland ones; avoid phrasing typical of mainland textbooks; if any historical date or allusion is uncertain, write "teacher to verify" rather than inventing it.

Below is a Secondary 3 student's Chinese narrative composition. Score Content, Expression and Structure out of 5 each, write two specific comments per criterion, then list three things this student should practise next.
Constraints: write in Traditional Chinese using Hong Kong conventions; do not rewrite the composition, give feedback only; quote the student's own sentences unchanged; point to specific paragraphs in your comments; if you are unsure about Hong Kong Chinese-subject marking conventions, mark it "teacher to verify".
[paste composition]

For wider comparison, read DeepSeek, Kimi, Qwen and GLM compared and the open versus closed cost and privacy trade-off.

The trade-off for schools

• If you need Chinese conversation data to stay entirely on the campus network, run Qwen3.8-27B or qwen2.5:14b on Ollama with Hong Kong usage fixed in prompt templates — do not reach for Max cloud for everyday planning.
• If you insist on the Qwen3.8-Max API, complete cross-border privacy assessment and parental notice first, and do not make personal-data marking the default route.
• If you need knowledge-base retrieval and Cantonese oral practice together, keep an OpenAI or Azure OpenAI key and run 14b and 27B in parallel for a term on 24GB hardware before settling on one size.