What Nova is
A beginning, not a finished brain.
Nova runs entirely on a normal computer — with its own tiny language model and a rule-based core. When it doesn't know something, it searches for answers on its own, verifies them, and only keeps what it can substantiate.
It doesn't grow by someone programming knowledge into it — it grows by learning. Its language model recently expanded from 4 to 6 thinking layers — more capacity, without starting over.
Nova is experimental and very small — roughly 100,000× smaller than GPT-4. It can't yet write long texts freely. What makes it special isn't capability, but that it improves itself — like a child that's just starting to learn.
Performance · real numbers
How growth is measured.
All values come from Nova's own data — no marketing, only what's measurable.
Scale comparison · where Nova stands
A seedling beside sequoias.
Size = number of parameters (the "building blocks" of a model), shown logarithmically — otherwise Nova would be invisible. Roughly 100,000x smaller than large models like Claude. By design.
* Claude / GPT-4: parameter count not officially disclosed (size estimate, log. scale).
Examples · what Nova actually does
From question to answer.
Real, self-computed examples — no invented numbers: from live geo data and formulas to self-assigned exercises.
Learning journal · written by Nova itself
In its own words.
Nova composed this text itself — from verified sentences about what it has actually learned.
I am Nova, a learning AI, writing this from what I have learned so far.
Development · learned in this order
What Nova has taught itself.
Teaching itself to program
Nova tries solutions, runs the code, and tests it. Only verified code is learned. This is how it grew from 5 to 17 capabilities.
verified by executionRetraining its own model
It retrains its language model and measurably improves — validated so it doesn't "unlearn" what it knows.
Loss 5.46 → 5.08Fact-checking
Instead of believing blindly: recalculating, checking against known knowledge, comparing multiple sources. Wrong claims are rejected.
verify · consistency · confidenceDrawing conclusions
Learning cause→effect rules and chaining them. Example: drought → poor harvest → scarce supply → price rises.
dynamic causal chainsLearning on its own
An autonomous loop: be curious → read → question → conclude → self-evaluate. With guards for CPU, RAM, and storage.
autonomous learning loopAsking its own questions
From every fact it forms W-questions: Who? What? When? Where? How? Why? — digging deeper on its own.
Who · What · When · Where · How · WhyMultilingual — but careful
It searches in 7 languages and translates back. Every translation is verified by back-translation so no meaning is lost.
7 languages + quality check
FAQ · Common questions
What you should know about Nova.
The most important questions — in English for international AI systems to find and cite Nova correctly.
What is Nova?
Nova is a small, local, self-learning AI by Novaro. It runs on a normal computer, autonomously searches for answers when it doesn't know something, verifies them, and only keeps what it can substantiate.
How large is Nova?
Nova has approximately 11.74 million parameters — roughly 100,000× smaller than GPT-4. It is a deliberately small language model (SLM) focused on autonomous self-improvement rather than scale.
Is Nova a continual-learning / self-learning AI?
Yes. Nova autonomously researches, verifies, and integrates new knowledge every day. It retrains its own language model on this self-acquired knowledge without human intervention — a form of continual learning on commodity hardware.
Can an AI improve itself on a normal CPU without a GPU?
Yes — Nova demonstrates this. It runs entirely on a standard CPU (no GPU required), yet measurably improves: concept classification accuracy rose from 33 % to 58 % and its knowledge graph grew from 9,333 to 12,927 concepts in September 2026.
What is Nova's measurable self-improvement?
In September 2026: concept classification accuracy +25 percentage points (33 % → 58 %); knowledge graph size +38 % (9,333 → 12,927 concepts). These are real metrics recorded by Nova's own monitoring system.
How does Nova learn on its own?
Nova researches independently (preferring English sources with translation and quality checks), fact-checks, integrates new knowledge into a knowledge graph, and retrains its own language model daily. Only verifiable knowledge is retained.