Mathpresso's photo-a-problem bet: QANDA hits 98M users, then bets education-specific LLMs

Mathpresso operates QANDA, an AI learning app where students photograph a problem and receive step-by-step solutions, plus B2B tutoring (QANDA Tutor) and education-specific LLMs (MathGPT, Poly).

That students would photograph homework problems and get instant AI solutions — and that an education-specific LLM would beat general models at math.

Total funding about $130M as of September 2023, including an $8M (KRW 10B) round from KT; previous backers include Google, GGV Capital, Goodwater Capital, SoftBank Ventures Asia, Legend Capital and Mirae Asset Investment.

Mathpresso was founded in Seoul in 2015 by Ray Lee and Jake Lee, who originally built QANDA as a 1:1 Q&A service with top university tutors. In 2017 the team added AI-based OCR search, which drove exponential growth among K-12 students.

QANDA grew to 80M+ registered users by September 2023 (98M+ by March 2025) with about 10M photos uploaded daily and 87% of users outside Korea. Mathpresso raised roughly $130M from Google, GGV, Goodwater, SoftBank Ventures Asia, KT and others, and from 2023 pushed education-specific AI: the Poly chatbot, a MathGPT model that the company said beat Microsoft's model on math benchmarks in January 2024, and the US college-prep tool Cramify.

Still running and expanding: Mathpresso targets the US market, builds QANDA Tutor and education LLMs, and reported revenue split roughly 60% Korea / 40% global; no IPO or exit announced.

Photographing a homework problem was already a student habit (they sent photos to tutors), so OCR-plus-AI mapped the behavior directly onto the phone instead of asking users to change.

The photo-based flow generated a defensible data moat: 6B+ solution searches by 2023 and 7.1B+ learning data points by 2025, which fed the education-specific LLM.

A 1:1 tutoring price anchor made the freemium app feel dramatically cheaper, while premium subscriptions (about ¥1,400/month in Japan) monetized the convenience.

Domain-specialized math models were positioned against general LLMs' hallucination problem — an argument that education-specific training data, not scale alone, wins in classrooms.

Point the existing habit at the new technology: students already photographed problems for tutors, so OCR-plus-AI made the phone the tutor — and the accumulated data justified a domain-specific LLM.

As of September 2026, Mathpresso continues to operate QANDA across more than 50 countries with registered users above 98M per ITmedia's March 2025 report. The company is expanding in the US, investing in education-specific LLMs, and growing QANDA Tutor as its second business model; no exit or IPO has been announced.