Chinese RedNote Model Scores Perfect Marks at IMO 2026 Math Olympiad

China's RedNote platform, more widely known as Xiaohongshu, has shown that its in-house language model can do more than recommend videos and beauty tips. The dots-note-3.0 model correctly solved all six problems of the 67th International Mathematical Olympiad, earning the maximum 42 points. It marks the first time a Chinese model has received an officially confirmed IMO gold medal.

What the result shows

The International Mathematical Olympiad is the most prestigious math competition for high school students, held since 1959. Its problems demand not just computational skill but, above all, original proofs and inventive constructions, which for years made it a tough proving ground for artificial intelligence. Language models could handle calculations but tended to falter when constructing complete, rigorous mathematical proofs.

The history of AI attempts at the IMO shows how quickly that changed. In 2024, Google's Gemini tackled the olympiad's problems for the first time and scored 28 points, a result on par with a silver medal. A year later, in 2025, Gemini Deep Think solved five of six problems in under 4.5 hours, earning gold. The dots-note-3.0 result from July 2026 goes a step further, with the model solving every single problem.

How the model works

According to reports from Chinese tech media, dots-note-3.0 is built on an agentic architecture that combines natural-language reasoning with Python code execution. The model uses a recursive self-critique mechanism that lets it continuously verify the correctness of its own arguments rather than relying on a single, one-shot answer.

Particular attention was drawn to the model's solution to the third problem, which dealt with combinatorial game theory. Instead of the standard graph-theory approach, the model applied inductive reasoning, capturing the structure of the problem in a way that math experts described as unusual even for human competitors.

Correct and elegant, with a compact structure and natural logic - Liu Hanzu, two-time gold medalist of the Chinese Mathematical Olympiad

Concise and direct, getting straight to the heart of the problem - Wang Qiantong, gold medalist of the Chinese Mathematical Olympiad

Chinese and American labs compete

The dots-note-3.0 result fits into a broader trend this summer, with lab after lab testing its models against this year's IMO problems, though not always through the formal evaluation process used for Google and OpenAI. Developers independently reported perfect scores from other models, including systems from Anthropic, OpenAI, and China's Moonshot AI, run on their own against the olympiad's problems. Still, official gold-medal confirmation from the IMO organizers, which Google received earlier and RedNote has now received, carries more weight than tests run independently by the developer community.

For RedNote, this is a reputational breakthrough. The company, known mainly for a social platform resembling Instagram that is popular with hundreds of millions of users in China and abroad, had until now developed AI models mostly for internal use, including the open-source dots.llm1 for language tasks and dots.ocr for document analysis. The IMO 2026 result shows that the company's research lab, referred to as hi lab, can compete in advanced mathematical reasoning with the industry's top players.

What comes next

Chinese tech media report that dots-note-3.0 is expected to be released as an open-source model, which would let the research community outside China independently verify and reproduce the result. Open-sourcing a model capable of a perfect IMO score would be unprecedented, since prior gold-medal models, from Gemini to OpenAI's systems, have remained closed commercial products.

For readers in Poland, this result is another signal that the lead in mathematical and scientific AI reasoning is no longer reserved for large Silicon Valley companies. Chinese labs, including DeepSeek and now RedNote, are increasingly delivering results at or near the top tier, accelerating the pace of competition to build models capable of supporting scientific research, engineering, and mathematical proof verification.