Leo 2B / 4B reasoning models
Small models, real intelligence
Compact AI tuned for reasoning — not just generation. Runs on classroom-grade hardware and aimeHUB locally, so every school gets capable AI without a data-centre bill.
Classroom AI
ai built for schools — not generic chatbots.
AI tutors and teaching assistants tuned to your country's curriculum, your teachers' ways of working and the standards your students are actually being assessed against.
Not
A generic chatbot.
Not
A consumer model rebadged for schools.
Made for
Your curriculum and your classroom.
Made for
Teachers and ministries to actually use.
Inside the lesson
Classroom AI delivered through the lesson — guided by the teacher, not bolted on as a side app.
What 'classroom AI' really means
Eight things that make ai work in real schools.
On your national curriculum
Explanations and examples are grounded in your country's syllabus, standards and grade-level outcomes.
Built around how teachers teach
Designed around lesson planning, scaffolding, marking and feedback — not just chat.
Thinks like the subject
Reasoning that follows how maths, science and language are actually taught — step by step.
Adapts in a sensible way
Pace and difficulty adjust to real progress against the curriculum — not just clicks or time-on-app.
Lives inside the lesson
Used by teachers and students inside the lesson flow — not as a separate consumer app.
Helps with assessment
Formative and summative assessment plugged into the AI loop and into ministry reporting.
Lines up with national standards
Outputs map directly to your country's frameworks, exam boards and curriculum codes.
Works in local languages
Strong support for national languages, dialects and bilingual classrooms.
Your curriculum, made usable
The national curriculum, powering every lesson.
What sets NDX apart from generic AI chatbots and eduAI tools: the whole platform is built around your country's curriculum — every objective, every assessment, every AI response traceable back to it.
The whole syllabus, in one place
Subjects, units and learning outcomes mapped to your national curriculum — auditable by ministry teams.
Aligned to national standards
Direct links to ministry frameworks, exam boards and grade progressions, refreshed each curriculum cycle.
AI that teaches the right things
Explanations and sequencing follow the curriculum, not a generic chatbot's best guess. Every response is traceable to a learning objective.
Assessments built in
Formative and summative assessment treated as part of the platform, with item-level evidence teachers can act on.
Track real progress
Follow how students are doing against the curriculum — mastery, not just clicks or time-on-app.
Adapted to your country
Local versions per country, region and language — without forking the platform or losing standards alignment.
Mapped to syllabusExam-board alignedLocalised by regionReviewed each cycle
Edu Intelligence
Silicon Valley can't build the ai education actually needs.
We didn't set out to build another AI product. We set out — from a deep understanding of how classrooms, teachers and ministries really work — to build a new category: educational intelligence. Small, specialised, local, and accountable to the people who have to teach with it.
Silicon Valley wasn't built for classrooms
Hyperscale AI labs optimise for advertising-scale inference, English-speaking power users and always-on bandwidth. None of those assumptions hold inside a Year 4 classroom in an emerging market — and no amount of API access changes that.
Education is a different problem class
Children, teachers and ministries don't need a more eloquent chatbot. They need reasoning aligned to a curriculum, pedagogy encoded in the model and outputs a teacher can defend in front of a class. That's a different engineering target.
Emerging markets are the real test
If an AI system can't run on a low-power device, on intermittent power, with no reliable internet, in a school that shares one classroom panel between two grades — it isn't built for the majority of the world's learners.
The shift we're driving
From cloud-scale generalists to local, specialised intelligence.
Personal and local computing, paired with small language models tuned for specific educational purposes, is how AI finally reaches every classroom — not just the well-connected ones.
From hyperscale to personal compute
Intelligence belongs next to the learner
We move computation from a distant data centre to the classroom itself — onto local hubs, panels and tablets. Latency disappears, data stays in the school, and lessons keep running when the connection doesn't.
From general models to specialised SLMs
Small models, trained for one job: to teach
Instead of a 400B-parameter generalist, we run 2–4B-parameter small language models distilled and tuned for specific educational tasks — explaining a concept, marking an answer, generating a diagram, scaffolding a question. Each one does its job better, cheaper, and offline.
From product to public infrastructure
Built with ministries, not sold over them
Educational intelligence isn't a SaaS subscription bolted onto a school. It's infrastructure — designed with ministries, aligned to national curricula, owned in-country, and engineered to outlast any single vendor.
This is what we mean by educational intelligence: AI that lives where learning happens, runs on the hardware schools actually own, and serves the curriculum a country actually teaches.
Why aime is built differently
Engineered from a deep understanding of what classrooms actually need.
Most AI in education is repurposed from chatbots — huge models, generic answers, always-online. aime is built the other way round: small efficient models, structured knowledge, and a delivery format made for real schools — including the ones with patchy internet and shared devices. We think differently and have created AI educational intelligence:
ThinkCache acceleration
Instant responses, no waiting
A custom thinking-cache reuses reasoning across steps, so tutors, lesson generators and assessments feel live — even on low-power devices in a busy classroom.
Kern + Loom agent stack
Reliable AI that just works
A lightweight agent framework and workflow engine coordinate lesson creation, marking and tutoring without the fragility of typical AI pipelines — ~78% more efficient than standard approaches.
ThinkBook knowledge layer
Truthful, structured answers
Knowledge is organised like a library, not scraped like a search index — drastically reducing hallucinations and keeping every answer grounded in the curriculum.
EduRule pedagogy engine
Teaching, not just answering
Lessons are shaped by encoded teaching methodology — explanations that build curiosity and intuition, not just correct text. AI that behaves like a good teacher.
Text-to-Diagram
Visual explanations, not generic art
Instead of slow, unpredictable image generation, aime produces clean educational diagrams and infographics — clear, accurate and built for understanding, not decoration.
aime Pack + Renderer
Lessons that travel anywhere
A purpose-built offline lesson format and runtime means AI-powered classes work without internet — and the same lesson plays on a tablet, a panel or a low-end device.
Together, these choices make aime usable where it matters most — in classrooms that can't depend on perfect connectivity, premium hardware or generic global AI.
Intelligence
How we think about ai in the classroom.
Most AI for education is a chatbot dressed for school. We took the problem apart and rebuilt it from the model up — a vertically integrated, AI-native learning system designed for real teachers, real curricula and real infrastructure constraints.
Small models, deep reasoning
We don't believe bigger always means better in a classroom. We distil reasoning from frontier systems into compact models that think well — not just generate fluently.
Knowledge, not scraped text
Education needs grounded answers. We structure knowledge like a library — hierarchical, linked, curriculum-aware — so the AI cites the syllabus, not the open internet.
Pedagogy is a first-class citizen
A correct answer isn't a good lesson. We encode teaching methodology — how great teachers build curiosity, scaffold ideas and check understanding — directly into the system.
Offline-first, always
The schools that need AI most have the least bandwidth. Every layer — models, agents, content, runtime — is engineered to work without an internet connection.
Architecture
Five layers, one coherent system.
Every layer exists to remove a real constraint — compute, connectivity, hallucination, pedagogy or distribution. They're engineered to work together, not bolted on.
Layer 01
Intelligence Layer
Reasoning
Reasoning models
Leo 2B / 4B
Multi-teacher distillation with reasoning-trace optimisation
Compact language models trained to reason, not just predict the next token. Distilled from advanced systems so a 2B-parameter model can hold its own against far larger ones — and run on classroom-grade hardware.
Performance layer
ThinkCache
Custom KV cache built for iterative thinking
A reasoning-aware cache that reuses intermediate thought across steps. Small models stop repeating work, agents respond in real-time, and a teacher's question doesn't stall the lesson.
Layer 02
Agent Infrastructure
Orchestration
Micro agent framework
Kern
Simplified dialect replacing heavy JSON tool-calling
A lightweight execution layer for small models. By replacing verbose JSON schemas with a compact interaction dialect, Kern cuts parsing overhead and pushes agent efficiency up by roughly 78%.
Workflow orchestration
Loom
Temporal-inspired engine, AI-native and lean
Long-running workflows, agent coordination and state — without the operational weight of enterprise orchestrators. Lesson generation, feedback loops and collaborative sessions all run through one resilient backbone.
Layer 03
Knowledge Layer
Grounding
Structured knowledge
ThinkBook
Library-inspired architecture beyond RAG
Traditional retrieval-augmented generation flattens knowledge into chunks. ThinkBook organises it the way curricula actually work — hierarchically, with concept-level links — cutting hallucinations and producing answers a teacher can trust.
Pedagogy engine
EduRule
Teaching-style encoding inspired by Derek Muller
EduRule transforms knowledge into teachable narratives. It encodes curiosity-driven explanation patterns, misconception checks and scaffolding so generated lessons feel like a good teacher — not a search result.
Layer 04
Content Generation
Explanation
Visual generation
Text-to-Diagram
LLM + parser pipeline, not diffusion
Diffusion models make pretty pictures; classrooms need clear diagrams. A structured pipeline turns concepts into clean, accurate, information-dense infographics — fast enough to drop into a live lesson.
Layer 05
Delivery Layer
Distribution
Offline content format
AIME Pack
Portable file format for AI-powered lessons
A purpose-built container for lessons, diagrams, agents and metadata. One file moves between server, tablet and panel — and a school with no internet still gets the full AI-powered experience.
Runtime engine
AIME Renderer
Unified view, edit and collaboration runtime
The environment teachers and students actually touch. Renders lessons, hosts agent workflows and supports live collaboration — the same surface across devices, online or off.
Why it matters
AI that works where education actually happens.
The result is a system that delivers high-quality, personalised learning at a fraction of the compute, infrastructure and connectivity cost of conventional AI — without compromising on pedagogy or trust.
~78%
more efficient agent execution vs. standard JSON tool-calling
2–4B
parameter models doing the work of much larger systems
0
internet required — full AI lessons run offline
1
unified runtime across server, tablet and classroom panel
AI as one part of one connected classroom platform.
AI doesn't arrive as a separate product. It comes connected with the classroom, the devices, the curriculum and the ministry's view across all of it.