Gemma 4 in Schools: A Local AI Model That Doesn't Need the Internet
Why it matters, who it’s for, and how to make it work
Not Another Cloud Tool. Something Different.
On April 2, 2026, Google DeepMind released Gemma 4 — and this time, something genuinely changed. Not just in benchmark numbers, though those are impressive too. What changed is more fundamental: a very good, multimodal language model now fits comfortably on a teacher’s laptop and requires no internet connection.
For three years, the conversation about AI in schools has revolved around ChatGPT and Gemini — cloud-based tools requiring accounts, collecting data, raising GDPR questions, and depending on a reliable WiFi connection during class. Gemma 4 starts from a different premise: you download the model once, install it locally, and it runs — on a school server, on a classroom laptop, in a computer lab without WiFi, on a Raspberry Pi in the library.
This article is not a step-by-step installation guide — that’s what documentation is for. It’s an attempt to answer the question that should come before installation: why would a school want a local language model at all, and how can it be used thoughtfully?
What this article covers: Gemma 4 is an open-source model released by Google DeepMind on April 2, 2026. The Apache 2.0 license means complete freedom of use — commercial and educational, no fees, no call limits. Data processed locally never leaves the school.
What Gemma 4 Is and What It Can Do
Gemma is Google DeepMind’s family of open-weight models — meaning you download the model weights and run them on your own hardware. Gemini (the one integrated with Google Workspace) stays in Google’s cloud. Gemma is its lighter, local cousin built on the same underlying technology.
Four Sizes, Four Scenarios
Gemma 4 comes in four variants. The “E” prefix on the smaller models stands for “effective parameters” — they’re more efficient than their parameter count alone would suggest:
The key number: The E4B model runs on a laptop with 8 GB RAM. A standard teacher’s ThinkPad or MacBook Air M1 with 8 GB unified memory is sufficient to run a full, multimodal language model offline. The 26B-A4B model — despite 26 billion total parameters — activates only 3.8 billion during each query thanks to its Mixture of Experts architecture, delivering speed comparable to a 4B model.
What Gemma 4 Can Actually Do
All Gemma 4 variants are multimodal from the ground up — not through optional plugins:
- Text: generation, summarisation, translation (140+ languages), proofreading, conversation, question answering
- Images: OCR (reading text from photos), chart analysis, document understanding, object detection, descriptions
- Audio (E2B/E4B): offline speech recognition — up to 30 seconds of audio, no API call required
- Video (26B/31B): video analysis up to 60 seconds, 1 frame per second
- Code: generation, completion, debugging — a dramatic quality leap over Gemma 3
- Reasoning: built-in step-by-step thinking mode, which the model can activate before responding
The context window is another surprising number: the E4B model handles 128,000 tokens — that’s the equivalent of an entire school novel with footnotes, loaded into a single prompt at once.
Why “Locally” Matters in a School
Before moving to use cases, it’s worth pausing on a question that rarely gets asked directly: why consider a local model at all, when Gemini and ChatGPT are available through a browser?
Argument 1: GDPR and Student Data
Every query sent to ChatGPT or Gemini reaches OpenAI’s or Google’s servers. If a teacher pastes fragments of student work, test results, or descriptions of difficult pastoral situations into a prompt — that data leaves the school and enters the data processing chain. Gemma running locally sends nothing outside the school network. Student data stays in school.
This isn’t paranoia — it’s a GDPR requirement when processing students’ personal data. The school as data controller should know where information about its pupils goes.
Argument 2: Availability Without Internet
Running an AI-assisted exercise in a lab where WiFi drops every other lesson? A field trip with tablets and no signal? A rural school with an asymmetric connection that buckles when 30 students are online simultaneously? A local model doesn’t know these problems. Downloaded once, it runs always.
Argument 3: Control and Customisation
A local model can be fine-tuned on your own school data. It can be configured to answer only questions aligned with the Year 7 physics curriculum. It can be embedded in a school application, an e-register system, or a library assistant. With ChatGPT you’re an API customer. With Gemma you own the model.
Argument 4: Cost at Scale
Gemma 4 is free. Apache 2.0 licence. No per-query fees, no monthly subscription, no call limits. A school serving 750 students making queries throughout the school year pays zero for model access — only for the hardware it runs on.
An honest caveat: Gemma 4 won’t replace Gemini or ChatGPT in every scenario. Cloud models have access to current information, better integration with external services, and — at their largest variants — higher quality on complex reasoning tasks. The choice between local and cloud isn’t ideology, it’s architecture: each use case is worth thinking through separately.
Gemma 4 vs. Gemini vs. ChatGPT — A Decision Map
Rather than a general comparison, here are concrete school scenarios with a recommendation for which tool fits better:
The conclusion is simple: Gemma doesn’t compete with Gemini — it complements it. Gemini works brilliantly as an assistant within Google Workspace. Gemma works brilliantly where data is sensitive, internet is unreliable, or the school wants its own configurable AI infrastructure.
Use Cases: What You Can Actually Do in the Classroom
1. Teacher’s Assistant for Instructional Materials
A physics teacher wants to generate 20 test questions at Year 7 level based on their own notes. They upload a PDF of a lesson script (OCR works even with handwritten notes), give the instruction, and 30 seconds later have a question bank — offline, no account, no materials sent externally.
The same applies to differentiation: the same material rewritten to three difficulty levels. Or translated into Ukrainian for a student with a migration background.
Example system prompt for a teacher’s assistant:
You are a teaching assistant in a Polish primary school.
Your tasks:
- Generate test questions and homework based on provided materials.
- Adapt difficulty level to the indicated year group (Years 5–8).
- Always respond in English unless the teacher requests otherwise.
- Never create content that could harm students.
- Do not comment on your technical limitations or architecture.2. Student Assistant — No Account, No Registration
Google AI Edge Gallery is a free Android application that runs Gemma 4 E4B locally on a phone. A student downloads the app and model (approx. 3 GB), and has a maths, language, or history tutor — no account required, no data collected, works on the underground without signal.
This matters: ChatGPT and Gemini require registration, which is legally problematic for students under 13 (GDPR, various national data protection laws). Gemma running locally has no account — it’s a tool like a calculator.
3. Audio Transcription and Analysis
The E2B and E4B models handle up to 30 seconds of audio without a network connection. Practical applications: a teacher records a student’s oral answer on a phone and gets a transcript for assessment; the school records a minute from a lesson and gets a summary; a student with dyslexia dictates instead of writing.
This isn’t a huge technical feat — but for a class with one child who has serious writing difficulties, it can change the quality of an entire year.
4. Image and Document Analysis (OCR)
Gemma 4 reads text from images with very high accuracy — including handwritten text. School scenarios: scan of a test → automatic feedback; photo of the whiteboard with notes → clean text for a summary; photo of a textbook exercise → step-by-step explanation.
Advanced variant: the 26B model running on a school server, accessible through a web interface — a teacher uploads an entire semester’s portfolio of student work (PDF) and asks: “Identify three areas where Matthew has shown the greatest progress and two where he needs support.”
5. Local School Chatbot (RAG)
RAG (Retrieval-Augmented Generation) is a technique combining a language model with your own knowledge base. The school loads documents into the database: the school handbook, timetable, extracurricular activities, parent FAQ. The model receives a question from a parent and searches for the answer in those documents — not on the internet, not in training data, only in what the school has provided.
The result: a chatbot on the school website that answers parent questions about duty hours, trips, and club sign-ups — 24 hours a day, in the right language, with accurate information, without making things up.
This already works: LeoAI on the PSP Piwoni school website in Szczecin is a deployed voice assistant built on ElevenLabs and RAG. Gemma 4 local is the next step — same concept, without dependency on external APIs, with full control over the knowledge base.
6. Tool for TechLab and Computing Lessons
Gemma 4 is excellent teaching material for students interested in AI. Instead of “using AI,” students can see how it works from the inside: download a model through Ollama, run it via terminal, write a simple Python script sending queries, compare responses with different system prompts.
This isn’t abstract. It’s a concrete technology whose weights can be downloaded, inspected, and fine-tuned. The difference between “I use AI” and “I understand what a language model is” starts exactly here.
Infrastructure: How to Set It Up in School
Scenario A: Teacher’s Laptop (Simplest Path)
Tool: Ollama — free, runs on macOS, Windows, and Linux, installs like any regular application.
# macOS / Linux — install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Download and run Gemma 4 E4B
ollama run gemma4:e4b
# Alternatively via LM Studio (GUI, recommended for teachers):
# 1. Download LM Studio from lmstudio.ai
# 2. Search for "gemma-4-e4b-it" in the model catalogue
# 3. Click Download → Load → ChatHardware requirements for E4B: 8 GB RAM (works, slower), 16 GB RAM (comfortable). MacBook Air M1/M2/M3 with 8 GB unified memory works very well thanks to Metal acceleration. Windows with an RTX or AMD RX card — full GPU support. Without a GPU: runs on CPU at approx. 2–5 tokens/second — slow, but usable.
Scenario B: School Server for the Whole School
The 26B-A4B model running on one machine with 16–32 GB RAM, accessible via the local network. All teachers use it through a browser — no installations on endpoints.
Scenario C: Student Phones and Tablets
Google AI Edge Gallery (Android app, APK installation — sideloading) runs the E2B or E4B model locally on a phone. Minimum requirements: 8 GB RAM and a Snapdragon 8 Gen 3, Dimensity 9300, or Apple A17 Pro chipset.
On iPhones with M-series chips (iPad Pro, MacBook) the E4B model runs through apps like LM Studio or Enchanted. This is a realistic scenario for BYOD schools.
Important limitation: Chromebooks — the most widespread hardware in many schools — do not natively support local LLM models. ChromeOS is a closed system without access to the GPU level required by llama.cpp or Ollama. Exception: Chromebooks with Linux mode (Crostini) on selected models can run Gemma E2B via CPU, but it will be slow.
Scenario D: Raspberry Pi in the School Library
It sounds like a novelty, but it makes sense: a Raspberry Pi 5 (4 GB RAM) runs the E2B model at approx. 5–10 tokens/second. Sufficient for a simple information kiosk, library assistant, or access terminal for students without their own devices.
# On Raspberry Pi 5 with Raspberry Pi OS (arm64)
curl -fsSL https://ollama.com/install.sh | sh
ollama run gemma4:e2b
# Web interface available on the local network on port 3000:
docker run -d -p 3000:8080 \
-e OLLAMA_BASE_URL=http://localhost:11434 \
--name open-webui \
ghcr.io/open-webui/open-webui:mainLimits and Questions That Shouldn’t Be Skipped
What Gemma 4 Does Poorly
Honest limitations:
- Knowledge has a cutoff date: January 2025. The model doesn’t know about events after that date. If asked about current news without being told, it won’t say it doesn’t know — it may simply be wrong. This matters in civics or contemporary history lessons.
- Hallucinations: like every LLM, Gemma can confidently assert things that are untrue. Students must understand that a language model is not an encyclopaedia — it’s a thinking tool, not a fact-verification system.
- No internet access is a feature, not a bug: but it means the model won’t suffice for tasks requiring current data.
- Quality on small hardware: the E4B model on 8 GB RAM is good, but it’s not Gemini Ultra. For complex, multi-step analysis, it’s better to use the 26B model or reach for the cloud.
Ethical Questions the School Should Ask
Deploying AI in school — whether local or cloud-based — requires answers to several questions that matter more than the technical architecture:
- Do students know they’re talking to AI? Transparency is fundamental — not just legally, but pedagogically.
- Where does support end and replacement begin? A model that writes a student’s essay for them doesn’t serve education. A model that helps structure a thought — it does.
- How do we assess AI-assisted work? The school needs a clear policy before a teacher encounters the first problem.
- Who has access to conversation logs? Even locally — if Open WebUI saves chat history, that is personal data. A procedure and legal basis are needed.
Good practice: Before launching Gemma in school, write one page of “AI Use Guidelines” for students and teachers. It doesn’t need to be perfect — it needs to exist. The conversation itself about how to use AI responsibly is a valuable lesson.
How to Start: A Roadmap
Resources to get started:
- Ollama — model installation and management
- LM Studio — GUI for macOS/Windows/Linux
- Open WebUI — web interface for the school server
- Google AI Edge Gallery — Android app with Gemma locally
- Hugging Face — Gemma 4 — model weights and technical documentation
Summary
Gemma 4 is not a revolution in education. The revolution will be what we do — or don’t do — with the tool.
A local language model solves concrete school problems: sensitive student data, lack of internet, cost at scale, dependency on external providers. It doesn’t solve questions about the meaning of AI in education — each school must ask those for itself.
Gemma 4 E4B on a teacher’s laptop, or a Mac Mini with the 26B model in the server room, is infrastructure that any school can deploy for the price of one new projector. The question isn’t “can we afford it?” The question is: “Do we know what we want to use it for?”
And that is the question worth starting with — before running the first ollama pull.
The infrastructure is real. The licence is free. The hardware is already in the server room.
The only thing standing between your school and a local AI that protects student data, works without internet, and costs nothing per query — is the decision to download the model and see what happens.
Every school has teachers who are ready for this. Very few have given them the infrastructure to try. That gap won’t last forever either.