GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

RedNote's dots-note-3.0 Scores 42/42 at IMO — First Perfect Score in AI History

RedNote, the Beijing company behind the social platform Xiaohongshu, has produced the first AI system to score a perfect 42/42 at the International Mathematical Olympiad. The model, dots-note-3.0, remains in beta. It is the smallest variant in the dots3 family — the larger jazz and aria versions are tuned for different use cases.

The prior frontier record was 35/42, set independently by Google DeepMind and OpenAI last year. At the 2025 IMO in Shanghai, only 7 of 666 human contestants matched a perfect 42. IMO judging requires written proofs, not final answers — a skipped case or unstated condition costs points even when the last line is correct.

Why This Score Matters More Than a Standard Benchmark

Most AI mathematics evaluations check only the final answer. A correct number obtained by flawed reasoning scores full credit. IMO does not work that way. Human graders read every line of the proof, and near-misses are penalized. That makes 42/42 here harder to inflate than a leaderboard number on AIME or AMC.

dots-note-3.0 read the original contest documents directly with no human reformatting. It ran an agentic loop that mixed natural-language reasoning with Python code it executed itself, tested its own draft proofs against counterexamples, and repaired failures before any output went to official graders. The system questioned its own assumptions repeatedly through the run; contest rules barred all external corrections or hints.

The Architecture of the Approach

RedNote has not released a technical paper, but described the agentic structure at a high level. The model pursued multiple proof strategies in parallel rather than committing to a single approach early. It explicitly hunted for counterexamples to its own intermediate lemmas — a step that distinguishes adversarial self-review from simple chain-of-thought prompting. Routes that led only to other open problems were discarded.

The approach echoes what a researcher using GPT-5.6 Sol described this week: a single operator solved six open Erdős problems in five days using a similar discipline — parallel search, adversarial checking, and killing dead-end routes before they consumed budget. Mathematical compute is now operating at a pace human patience cannot match.

What Is Not Yet Known

RedNote has committed to open-sourcing dots-note-3.0 but has not attached a timeline. The model has not appeared on standard independent benchmarks such as AIME 2025, LiveBench, or any HLE tier. The claim rests on competition results certified by IMO judges, which is a credible verification channel — but independent academic analysis of the model’s architecture and failure modes has not been published.

The company also operates under Chinese data-sharing regulations that differ from those governing Western lab releases. Whether the open-source commitment extends to weights or only to architecture details is not yet specified.

Key Numbers

  • IMO 2026 perfect score: 42/42
  • Previous AI best: 35/42 (Google DeepMind and OpenAI, 2025)
  • Human contestants at 42/42 in Shanghai: 7 of 666
  • Model: dots-note-3.0 (smallest in dots3 family; jazz and aria are larger)
  • Status: beta, open-source promised with no timeline