DeepSeek (V4.1-Flash): the cheapest open-weight flagship, and the two realistic routes for a Hong Kong school

Procurement slides often label DeepSeek "MIT open weights at $0.15 per 1M tokens" — yet a school actually chooses between two routes: Ollama on campus with a distilled build and no data leaving the network, or the official mainland API as a cross-border transfer that needs a privacy assessment first; downloadable and runnable flagship are not the same thing.

DeepSeek was founded in China in 2023, and DeepSeek-R1 in January 2025 reset the assumption that frontier reasoning required frontier training budgets. For a Hong Kong school the value is MIT licensing and rock-bottom API list price; the reality is flagship self-hosting limits, peak-hour doubling, and Traditional Hong Kong wording that still needs teacher review.

What the company is actually building

DeepSeek's technical direction stays on one theme: buying down inference cost through MoE architecture and KV cache compression. The current DeepSeek-V4.1-Flash (10 September 2026) is a 552B-parameter MoE activating 8B at prefill and 16B at decode, with native image input, a 1M-token context window, 384K output, and weights under MIT — the most attractive permissive licence for a sponsoring body that wants full deployment control.

On the same trunk sit DeepSeek-V4-Pro and V4-Flash (16 August 2026), the generation that introduced peak and off-peak pricing; DeepSeek-R1 (20 January 2025), still the name most often raised in school self-hosting discussions; and DeepSeek-V3 (26 December 2024), which brought MoE into mainstream conversation. For the classroom, the Ollama R1 distillations from 1.5B to 70B matter more — those are the sizes school hardware can actually run.

Model line-up at a glance

The $0.15 / $0.60 rate is off-peak only; a Hong Kong school's daytime usage may not map to that window. Annual estimates should model peak doubling or a mid-range figure, not lowest rate times volume.

Where the money actually goes

Cloud API: list price sits at the bottom of the major-provider range, suited to high-volume, moderate-difficulty work after a cross-border privacy assessment — rewriting public English material at different levels, compiling topic background from open sources, drafting practice questions. Marking that involves student names, classes or written work is better left on an assessed native provider.

On-premises: hardware is capital spend, but a distillation is noticeably weaker than the flagship. An IT coordinator should run a mid-sized Ollama build on an existing workstation first, measuring Traditional Chinese output quality and latency, before buying. On licensing, V4.1-Flash is MIT; distilled builds need Apache 2.0 or Llama checks one by one.

Hidden cost: DeepSeek supplies no embedding or speech models. Even with self-hosted chat, the knowledge-base vector index and oral practice still need OpenAI or Azure OpenAI. That is not picking the wrong vendor — language models and multimodal infrastructure were never the same product.

Putting $0.15 on a principal's slide without "peak doubling, cross-border API, flagship won't run here" is quoting rent without the deposit — procurement papers need all three layers.

What role it plays inside Edor.ai

DeepSeek is not a native provider. The native routes are OpenAI, Azure OpenAI, Anthropic, Poe and local Ollama; there is no dedicated DeepSeek connector.

Route one: local Ollama (recommended)

A distilled or quantised build on a school server keeps inference inside the campus network. Four-layer gating, teacher monitoring and ai_audit_logs operate normally. Combined with tutor mode, this is the most controllable option — after teachers test distilled Chinese feedback quality and decide which subjects it serves.

Route two: official API (assessment required)

Servers sit in mainland China, so requests containing student personal data are cross-border transfers. Complete an assessment under the Data Protection Principles: data categories, purpose, protection at destination, alternatives, signed and filed by management. We would not enable student-facing use before that step.

On either route, embedding and speech still need OpenAI or Azure OpenAI. If the goal is data residency, read how the platform handles data and security and the 27B-class self-hosting path under Alibaba Qwen.

Where it fits and where it does not

Fits: high-volume teacher paperwork with no student personal data after privacy assessment; student hints on a campus Ollama distillation in tutor mode after Chinese quality testing; IT proof-of-concept on existing hardware before a procurement case.

Does not fit: student work on the cloud API before assessment; selling a 14B distillation with flagship benchmarks; annual budgets built only on off-peak $0.15; expecting DeepSeek alone to run knowledge base and Cantonese oral practice; adopting mainland phrasing without teacher review.

After hours, an IT coordinator tells the principal: "We ran a 14B distillation on Ollama — about eight seconds a reply, General Studies piloted fine for a month, but Chinese said the wording still needs checking." That is a sensible pilot rhythm.

Copy these prompts to test:

You are a Hong Kong primary General Studies teacher. Design a 25-minute classroom activity on "weather and climate" for Primary 4, using only simple equipment a school would already have.
Output: three activity objectives, a procedure written in short sentences pupils can follow, an observation record table, and two extension questions.
Constraints: write in Hong Kong English conventions and use the vocabulary Hong Kong schools use, for example "worksheet"; do not give away answers; for any specific Hong Kong temperature or rainfall figure, write "teacher to verify" and say which Hong Kong Observatory resource should be consulted.

Below is a Secondary 2 Chinese-language reading passage and five draft comprehension questions. Please review the quality of the questions.
Output: for each question, say which skill it tests (retrieval, inference or evaluation), identify any question that can be answered without reading the passage, and propose one rewritten version of each question that needs improvement.
Constraints: write your commentary in Hong Kong Chinese conventions; do not rewrite the passage; do not introduce material beyond the passage; if you are unsure whether a word suits the Hong Kong primary or junior secondary level, write "teacher to verify".
[paste passage and draft questions]

For a side-by-side view of Chinese vendors, 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 data to stay entirely on the campus network, run a distilled or quantised build on Ollama and test Traditional Chinese feedback before student access — do not assume flagship capability transfers to 14B.
• If you call the official API, complete and file a cross-border privacy assessment first, and leave tasks involving student names and written work on an assessed native provider.
• If you need knowledge-base retrieval and oral practice, keep an OpenAI or Azure OpenAI key and model peak doubling in cost estimates, not off-peak $0.15 / $0.60 alone.