Education is not a free training set.
Every evening teachers paste tests, pupil work and parent mails into chatbots run by American companies. PrivateGPT is the AI that belongs to the institution: it runs in the Netherlands, it knows your own materials and it learns nothing from what is typed into it.
Write a mail to the parents of year 9 about the new phone rules. Use our own protocol.
From Monday 6 October phones stay in the locker for the whole day, breaks included. That is the rule from our protocol Phone free learning (handbook, p. 14). What this means for your child: you can reach them through reception, and…
PrivateGPT in a nutshell
See in under a minute how institutions use AI without giving away their data.
What happens today
On an average Tuesday, without anyone doing anything wrong.
22.01 · at home
Mark this test and add grades.
test_ch3_term2.docx
Processor unknown
At 22.00 a teacher pastes a test into a free chatbot. At 22.01 it sits on a server the institution has no contract with.
23.40 · bedroom
essay_final.docx
Private account · invisible to the institution
A pupil has his essay written for him. Nobody inside the institution sees it: the account is his, not the organisation's.
08.15 · Monday meeting
“Can we arrange something with AI for the departments?”
Request #248 · IT support
No overview of what is already running
IT is asked to arrange something with AI, but has no view of what is already running.
Banning it has been tried. That did not work. What is left is choosing who sets the rules.
The same question. Two answers.
On the left a free chatbot. On the right PrivateGPT, set up by the institution and fed with your own rubric.
Free chatbot
This is the essay by Noor from 2B. Give feedback based on our rubric.
essay_noor_2B.docx
Nice essay! Here are some tips:
1. Use more adjectives.
2. Watch the spelling of 'their'.
3. Make your conclusion longer.
Would you like me to rewrite the essay for you?
Which rubric? The chatbot does not know it. The essay and Noor's name now sit on a server outside the EU, where training on them is the default.
PrivateGPT · Montessori Lyceum Zuid
This is the essay by Noor from 2B. Give feedback based on our rubric.
essay_noor_2B.docx
Based on 'Writing rubric · lower years' (institution curriculum)
Structure: sufficient. The introduction asks a question, but the ending does not answer it yet.
Language: good. Varied sentences. Two tense mistakes in the middle paragraph.
Next step for Noor. Rewrite the ending in three sentences that point back to your question.
Source: Lower years language curriculum · rubric
Stays inside the institution's own space in the Netherlands. Nobody trains on it. The rubric comes from your own curriculum.
An AI that does not know your materials is a worse Google.
PrivateGPT works from the sources the institution adds itself: teaching methods, curricula, protocols, the handbook. That is why a lesson plan follows your curriculum and a parent letter sounds like yours.
For teachers, pupils and leaders, from primary class to lecture hall. Many national AI projects only serve higher education. PrivateGPT is built for the whole education system, primary and secondary very much included.
Sources from the institution
Added by the leadership team
Which learning goals belong to theme 5, and write three test questions for them.
Learning goals (from your curriculum): the pupil explains why the Republic grew rich, and names two darker sides of that wealth.
1. Why did so many ships sail to Amsterdam in the 17th century?
2. What did the Dutch East India Company have to do with slavery?
3. Name one thing that was golden and one thing that was not about the Golden Age.
Source: History year 6 · theme 5
The institution is in control
Who may do what, which sources are used and how long conversations are kept: the institution sets that. Not us, and not the company behind the model.
Rules for pupils and teachers
Applies to every year group
Off until the review in November
Department leads approve
Follows the institution's retention policy
Sources from the institution · 4
Handbook 2026–2027
Phone free learning protocol
Language curriculum lower years
Assessment policy upper years
Sign in
Microsoft Entra ID · institution accounts No separate passwords
Audit log · today
Microsoft Entra ID · no separate passwords
European and open models. For education work the difference is a rounding error.
PrivateGPT runs on GLM-5.3, Mistral Large 3, Qwen3-VL and Ministral 3.
The American top models win at maths olympiads and writing software. A parent letter, feedback against a rubric or a lesson at three levels does not ask for that. So PrivateGPT runs on European models and on open models we host ourselves, all on servers in the Netherlands.
European and open models, hosted in the Netherlands
Index score is the Artificial Analysis Intelligence Index (v4.3.2, September 2026), an independent composite of ten benchmarks. The highest score any model currently reaches is 53. A model scoring 45 sits just below the very best in the world.
Model size is shown in billions of parameters. Memory is the approximate number of pages that fit in one conversation.
90% of the daily work
Current world maximum 53
GLM-5.3 · 753B~1,500 pages of memory · no image input
Tests, rubrics, qualification files
Same engine. Deep Think lets it reason for longer before answering.
GLM-5.3, high reasoning effort · 753B~1,500 pages of memory · no image input
Parent letters, second language, translation
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
Mistral Large 3 · 675B~390 pages of memory · reads images
Worksheets, whiteboards, handwriting
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
Qwen3-VL 235B · 235B~390 pages of memory · reads images
Short questions, pupil chat
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
Ministral 3 14B · 14B~390 pages of memory · no image input
Versus the commercial platforms
Index score is the Artificial Analysis Intelligence Index (v4.3.2, September 2026), an independent composite of ten benchmarks. The highest score any model currently reaches is 53. A model scoring 45 sits just below the very best in the world.
GLM-5.3 · Mistral Large 3 · Qwen3-VL · Ministral 3
Current world maximum 53. A score of 45 sits just below the very best.
BrainBite
OpenAI
Anthropic
The index is weighted towards reasoning and coding. A model built for fluent writing scores lower on it without being worse at writing.
The difference that counts is not where a model was made, but where it runs. For an open model we put the weights on our own servers in the Netherlands, so not a word goes back to the maker. With Claude or GPT your text goes to American servers by definition.
The institution picks per group which model is on: pupils on the standard model, a department on the heavy one. Everything runs on servers in the Netherlands and no model learns from what is typed into it. Does a board want European makers only? Then we switch the open models off.
Everything else you simply expect is in there
In the conversation
Upload files: PDF, Word, Excel, photos
Understands images
A whole method in one conversation
Tables and formulas
For the team
Dutch and English
Share templates with the department
Export conversations
For administrators
Sign in with institution accounts
Roles and rights
Audit log
What the platform can do
Pick a task and see exactly how to start using it in class.
These capabilities are available, but they are not on by default.
Standard
Handles most everyday education work.
- How to use it
- Type your question and add a file if needed.
- Classroom example
- Create a lesson outline about fractions for year 7.
Engine and index: GLM-5.3 · 45 of a current maximum of 53
Deep Think
Handles questions that need more reasoning steps.
- How to use it
- Choose Deep Think and include the assessment criteria.
- Classroom example
- Build a rubric and check that every learning goal is covered.
Engine and index: GLM-5.3 · same engine as Standard
Multilingual
Writes and translates naturally across languages.
- How to use it
- Paste your text and say who it is for and which language you need.
- Classroom example
- Write a short parent letter in clear English and Dutch.
Engine and index: Mistral Large 3
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
Vision
Understands photos, worksheets and handwriting.
- How to use it
- Upload a photo of the worksheet and ask.
- Classroom example
- Review a completed worksheet and give three useful feedback points.
Engine and index: Qwen3-VL 235B
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
Quick
Answers short, clear questions quickly.
- How to use it
- Type a short question without extra background.
- Classroom example
- Create five warm-up questions for a lesson about climate.
Engine and index: Ministral 3 14B
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
Help
Helps you find the right place or feature.
- How to use it
- Describe what you want to do in the platform.
- Classroom example
- Ask where to find a shared template.
Engine and index: Platform help · not applicable to this index
Document search
Finds answers in your institution’s approved sources.
- How to use it
- Paste the PDF, then ask about it.
- Classroom example
- Check the assessment policy to see when a resit is allowed.
Engine and index: Search and ranking layer · not applicable to this index
Speech to text
Turns spoken words into text.
- How to use it
- Press the microphone and speak your observations.
- Classroom example
- Dictate quick lesson notes for your report.
Engine and index: Speech service · not applicable to this index
Live captions
Shows spoken words as text while someone talks.
- How to use it
- Switch on live captions before the conversation starts.
- Classroom example
- Let a pupil follow an explanation with text on screen.
Engine and index: Speech service · not applicable to this index
Read aloud
Reads an answer aloud at a steady pace.
- How to use it
- Open an answer and choose read aloud.
- Classroom example
- Let a pupil hear the explanation while reading along.
Engine and index: Speech service · not applicable to this index
OCR
Pulls text from scans, photos and handwriting.
- How to use it
- Upload a scan or photo and ask it to read the text.
- Classroom example
- Make an old worksheet searchable.
Engine and index: DeepSeek-OCR · not applicable to this index
Kimi K3
An additional open model for language and reasoning tasks.
- How to use it
- Ask an administrator to enable it for a test group.
- Classroom example
- Compare two explanations for the same lesson.
Engine and index: Kimi K3
Current world maximum 53
GLM-5.3-Flash
A faster variant for short tasks.
- How to use it
- Ask an administrator to enable Flash for quick questions.
- Classroom example
- Create ten short quiz questions in one go.
Engine and index: GLM-5.3-Flash
Current world maximum 53
Qwen3Guard safety screening
Checks prompts and answers for safety risks.
- How to use it
- Ask an administrator to switch on extra safety screening.
- Classroom example
- Flag risky pupil prompts for a teacher to review.
Engine and index: Qwen3Guard · not applicable to this index
Z-Image-Turbo illustrations
Creates simple illustrations for lesson materials.
- How to use it
- Describe the image you want beside your explanation.
- Classroom example
- Create a calm illustration of the water cycle.
Engine and index: Z-Image-Turbo · not applicable to this index
Mistral Medium 3.5
An additional European model for general tasks.
- How to use it
- Ask an administrator to enable it for a test group.
- Classroom example
- Compare feedback on the same writing task.
Engine and index: Mistral Medium 3.5
This index measures reasoning and coding at PhD level, not how well a letter reads in Dutch. This model was chosen for language quality.
In writing
Dutch data centres
Not under the American Cloud Act. Data processing agreement for education, DPIA template included.
Nobody trains on your data
Not us, not the model supplier. Fixed in the contract.
You can stop at any time
Cancel monthly. All data comes with you, plus a certificate of destruction.
Training included
We train the team, set up the space and review together what works in class.
Standards and certifications
Certificates, scope and data processing agreement on request
- ISO 27001
- ISO 27701
- NIS2
- GDPR
- EU AI Act
We work with, among others
- TU Delft
- Radboud Universiteit
- TU Eindhoven
- Dutch EdTech
Scan
Take the scan, get advice for your institution
Seven short questions about your situation. You get advice on how PrivateGPT fits your organisation straight away. No email address needed.
What kind of education do you work in?
So we know which ages and rules to take into account.
Free pilot, two weeks
Up to 20 teachers or staff, with training and guidance. Limited places.
Per teacher or staff member per month
Excluding VAT, cancel monthly. Pupils and students through managed capacity, without separate licences.
Enterprise education
For boards and large institutions. Own infrastructure, on-prem included.
Want to see what it does with your own handbook?
We call within one working day. Bring a protocol or a teaching method and we let PrivateGPT work with it live. Thirty minutes, online.
No obligations, no installation
With your own documents as the example
We also talk through the free two week pilot