Xiaohongshu's dots Model Wins Perfect Gold at IMO 2026
Core Highlights
Xiaohongshu's dots team brought its internal dots-note 3.0 to the 67th International Mathematical Olympiad (IMO 2026), scoring full marks on all six problems for a perfect 42/42 gold medal. Worldwide, only seven human contestants achieved the same result. This means domestic large models have reached the very forefront of the world in formal reasoning and complex mathematical problem solving, a domain that has long been treated as one of the hardest tests of machine intelligence. The achievement is notable not only for the score but for the method, which departs from the heavy formal-tooling tradition that has dominated recent competition entries.
The result also matters because it comes from a consumer-internet company rather than a dedicated research lab, suggesting that advanced reasoning is spreading beyond the usual suspects. For the broader Chinese AI ecosystem, a perfect IMO score by a homegrown model is a morale and talent signal as much as a technical one, demonstrating that the country's engineering base can compete at the very top of the discipline.
What Happened
Unlike many competing systems that rely on formal languages such as Lean or Coq to assist proofs, dots-note 3.0 does not depend on formal languages. Instead, it reads raw LaTeX problems directly and completes the solution end to end through "recursive self-critique." In other words, after generating an answer, the model repeatedly scrutinizes its own reasoning steps, finds and corrects mistakes, and produces a complete and correct proof. This closed loop of "posing the problem itself, verifying itself, and correcting itself" is the key to its stable perfect score on highly difficult tasks.
dots-note 3.0 is the lightest version in the dots3 family, yet it exhibits a depth of reasoning disproportionate to its size, suggesting that careful architecture can matter more than raw scale for narrowly defined, logically intense tasks that reward rigor over breadth. The fact that a compact model can beat most human contestants on the planet is a striking data point in the ongoing debate about whether bigger is always better.
Technical Details
"Recursive self-critique" can be understood as the model constantly backtracking within its reasoning chain: each time it writes a sub-conclusion, it tries to attack it from the perspectives of counterexamples, boundary conditions and logical consistency, and rewrites the relevant passage if a flaw is found. Compared with outputting an answer in one shot, this multi-round self-dialogue significantly reduces the probability of hallucination and skipped steps, which are the two failure modes that usually defeat large models on olympiad-style problems.
Notably, this version does not pursue parameter scale, but trades it for real-time deployability through an efficient architecture. The team expects to open-source the model so the community can continue refining it, which may accelerate external research into lightweight reasoning systems and lower the cost of experimentation for smaller labs that cannot afford to train frontier-scale models from scratch. Opening the weights turns a one-off result into a shared platform.
Comparison with Competitors
In top-tier math competitions like the IMO, systems that previously performed well mostly relied on large-scale formal verification toolchains, which carry high deployment costs and weak generality. dots-note 3.0 chooses the lightweight route of "native LaTeX plus self-critique," requiring no accompanying formal environment and lowering the barrier to use. Compared with general-purpose large models with hundreds of billions of parameters, it achieves stronger specialized reasoning with a smaller footprint, validating the feasibility of the "small but specialized" route in mathematical scenarios.
Where rivals needed bespoke proof assistants and long compute budgets, dots-note 3.0 suggests that a disciplined self-checking loop can substitute for much of that machinery without sacrificing correctness. That lesson could ripple outward: if narrow, well-architected models can match giants on hard tasks, the economics of building AI shift in favor of clever design over brute force.
Industry Impact and Use Cases
Simply put, a perfect IMO score is not an endpoint but a milestone marking the maturity of large-model reasoning. This capability can transfer to scenarios with extremely high demands for logical rigor, such as automated theorem proving, scientific computing assistance, program verification and financial modeling. If dots-note 3.0 is open-sourced as expected, it will significantly lower the threshold for using advanced reasoning models in research and education, and may also push more teams to explore the "recursive self-critique" paradigm.
For Chinese AI, this proves that we already possess the strength to compete head to head with the world's top laboratories in frontier reasoning architectures, and it offers a credible domestic alternative to foreign reasoning systems that may face export or access restrictions in the years ahead. The deeper message is that reasoning, not just fluency, is now a battleground where local teams can lead.
Beyond the medal table, the result reframes how Chinese internet companies are investing in fundamental AI research. A platform company known for lifestyle content has shown it can field a system that beats nearly all human mathematicians, a sign that talent and ambition are no longer concentrated in a handful of Western labs, and that the next breakthroughs may come from unexpected corners of the industry.