Small language models for school work

During a tea break at a staff workshop a teacher lifts her cup and asks: "If the flagship costs only a few hundred dollars a month, why bother with 8B models? Why not always use the biggest?" The IT coordinator answers with a question: "Would you charter a fifty-seat coach to take thirty pupils to the swimming pool?"

This closes module 3. After reading model cards and file formats, the remaining question is how large a model your school's daily work actually needs.

---

The trip decides the vehicle

School travel usually works two ways. On normal days a minibus follows a fixed route with just enough seats and a driver who knows every turn. For a rare whole-school outing you charter a coach with more seats, higher cost and advance booking.

Large and small models relate in the same way. The question was never which vehicle costs more, but how many passengers, how far, and whether you stay overnight. Using a flagship for every school task is like chartering a coach for the daily run. It can be done, but it is expensive, slower, and sends sensitive data off campus when it need not leave.

Specialists trained on curated textbooks

A common misunderstanding is that small models are large ones with layers removed. The main reason they have improved so much recently is the quality of the training material.

Take Microsoft's Phi family. Its design idea is to train on large volumes of carefully filtered and synthesised "textbook-like" material rather than a copy of the whole web. The result is a 14B parameter model that approaches or beats far larger rivals on mathematics and reasoning benchmarks. In education terms it is a pupil taught from excellent textbooks rather than one who read widely but indiscriminately.

Google's Gemma family takes a different route, turning the research behind the flagship Gemini models into downloadable open weights deliberately built to run on ordinary hardware. Recent versions accept text, image and audio input and cover more than 140 languages with a long context window.

The 7B to 9B models from Meta, Alibaba Qwen and Mistral are what schools most often deploy locally, because Chinese support and community resources are comparatively mature there. Current line-ups are in the model database.

One caution: every model card states its language support. The Phi family's card, for example, says explicitly that it targets English and is not designed for multilingual use, which is a significant warning for Chinese-medium school work. How to read those pages is in reading a model card and what GGUF and Safetensors mean.

Most daily work fits a minibus

From what we see in schools, the following are practically indistinguishable on an 8B class model:

• turning a transcript of a meeting into a list of decisions;
• converting a teacher's paragraph of instructions into a worksheet table;
• rewriting one comprehension passage at Primary 4, 5 and 6 difficulty;
• classifying a hundred open-ended survey answers from pupils;
• checking whether a circular has omitted the date, venue and reply deadline;
• building a vocabulary list from an English article with difficulty labels.

What they share is this: the answer comes mainly from the material you supplied, not from the model's memory. That is where small models shine: data stays in school, prompts are clear, and output formats are fixed.

Four occasions that still need a coach

• Problems needing several steps of reasoning, such as a Secondary 5 mathematics proof or a full physics solution.
• Finding internal contradictions inside a long document.
• Outward-facing writing where tone matters enormously, such as replying to a complaint from a parent.
• Complex judgements that combine images and text.

What these need is depth of reasoning rather than breadth of knowledge, which is covered in multimodality, reasoning models and embodied AI.

Small models handle volume; flagships handle judgement. Mixing both into one pipeline is often slower, costlier, and sends raw sensitive text to the cloud when it did not need to go.

Two layers for nine hundred survey answers

A secondary school's careers team had a survey to process: nine hundred pupils, three open-ended questions each, about 2,700 pieces of text. Sorting that by hand would have taken one teacher a full week.

They worked in two layers. The first layer used an 8B model running on the school server to file each answer under further study, employment, emotional support, extracurricular activities or other, and to pull out keywords. That handled all 2,700 answers, and the pupils' own words never left the campus network for a second.

Only the second layer sent the de-identified statistical summary to a cloud flagship model to draft an analysis for senior management, because that report needed careful tone and inferences across categories.

The result: one teacher finished a week's work in two free periods, and the most sensitive raw text stayed in school throughout. Hardware costs fell within the Hong Kong EDB AI-in-education grant of up to HK$500,000 per school. The point was not "buy the biggest machine" but "buy one that comfortably runs a quantised 8B model".

Common misreadings

• The claim: Price or parameter count tells you how hard a task is. What actually happens: Ask first whether the answer lies in the material you supplied. Self-check: if you turned off the model's memory and gave a human assistant only the source text, could they still do it?
• The claim: Fine-tuning replaces a knowledge base for school facts. What actually happens: Rules, timetables and this year's circulars should be retrieved, not baked into weights. Self-check: see the school knowledge base and retrieval-augmented generation.
• The claim: Sensitive material should always go to the flagship because it is "more accurate". What actually happens: Compositions, counselling notes and SEN-related documents should stay on an in-school 8B model whenever possible. Self-check: if this material leaked, would the consequence be acceptable?

Try it once

Time: about twelve minutes. Materials: an 8B class model (local or cloud) and a flagship if you have both; the first half still works with only one.

Task: run the prompt below on both models and record the outputs:

Below are five pupils' answers to "what extracurricular activity would you most like the school to add". Sort them into no more than three groups, name each group, and list the answer numbers in it.
Classify only from the answers I have supplied. Do not add your own ideas and do not evaluate the pupils.
1. A photography club, because I like taking pictures.
2. More ball game competitions.
3. I want to learn cooking.
4. A badminton team.
5. I want to learn video editing.

Then change the prompt to "analyse how adding these activities would affect Other Learning Experiences under the current Hong Kong senior secondary curriculum" and run both again.

Hand in: four lines of notes: (1) on the first prompt, how large was the gap; (2) on the second prompt, did the gap appear immediately; (3) which task you would permanently assign to an in-school 8B model; (4) which task you would escalate to a flagship and why in one line.

Done right: you can state the boundary between "material-in-the-prompt tasks" and "reasoning tasks", and name one concrete school job your panel would switch to a small model next week.

That closes module 3. The next module begins with prompt engineering in one lesson. To revisit file formats, return to reading a model card and what GGUF and Safetensors mean. To see why small models are cheap to fine-tune, read LoRA and QLoRA on one GPU.