IOAI brings teams of up to four secondary students to an Individual Contest held over two six-hour days on identical GPU machines. An at-home round comes first, and the event also includes a Team Challenge and the GAITE contest. The top half of the individual field receives medals in a 1:2:3 gold, silver, and bronze ratio.
IOAI 2026 was held in Astana, Kazakhstan from 2 to 8 August. The 2027 event runs from 4 to 10 July in Singapore.
How a IOAI cycle runs, from registration to the contest itself.
Around 790,000 students sit national qualifications each year. Roughly 0.038% reach IOAI.
AOAI was held in Sousse, Tunisia from 9 to 12 April 2026, NAOAI ran across North America from 4 to 5 April, and APOAI took place in China from 13 to 20 June. EOAI and AMOAI follow in 2027.
Tasks are released early so teams can prepare. Day 1 on site extends them, so this round is not optional practice.
The third edition, in Kazakhstan, with 106 of 131 accredited countries registered.
The fourth edition, hosted by AI Singapore and NTU Singapore.
How the contest works and what the rules allow.
The Individual Contest is the only one that awards medals. The Team Challenge runs alongside it with a different format each year - AI image and video generation in 2024, programming Galbot humanoid robots for a warehouse task in 2025. GAITE is a simplified, hint-enabled track introduced in 2025 to widen access for underrepresented regions.
An at-home practice round first, then on site a roughly 90-minute practice session followed by two contest days of six hours each. Day 1 extends the at-home tasks; day 2 introduces entirely new ones with leaner compute budgets and unfamiliar domains, deliberately testing generalisation rather than preparation.
Every contestant uses the same local machine with at least 24 GB of GPU memory. PyTorch and scikit-learn are installed; TensorFlow and Keras are not. In 2025, contestants could reach a small set of approved documentation and a translation tool. The 2026 rules remove that internet access and leave only the provided assistant.
Contestants get GPT-4o-mini with a token cap on each query. Use it for documentation lookups and debugging. It is deliberately limited and is not meant to solve a task for you.
Each task is normalised to a score from 0 to 100. Your raw metric is mapped between a baseline minimum and the higher of 90% of the Scientific Committee's own solution or the best contestant submission. A modest improvement over the baseline earns real points, while chasing the last fraction of accuracy often costs too much time.
The top 50% of individual contestants receive medals, distributed gold:silver:bronze in a 1:2:3 ratio. The top three individuals receive honorary trophies, as do the top three teams in the Team Challenge.
Each division follows the official IOAI syllabus and its problem style.
Python, the scientific stack, and the linear algebra and probability the contest assumes you already have.
Classical learning and the evaluation habits that keep a normalised score above the baseline.
PyTorch training loops, transformers, and the fine-tuning techniques that fit inside a six-hour window.
Text and speech: tokenisation, BERT-family encoders, language modelling, and pre-trained audio models.
Convolutional and transformer vision models, CLIP, GANs, and diffusion, including direct model editing.
How to work on a locked-down machine with a capped assistant, a 24 GB GPU, and a strict clock.
Students reach IOAI through their national olympiad.
Supported by
An unofficial study guide, part of OpenIntelligence. Problem materials belong to IOAI.
© 2026 OpenIntelligence.