simplified-reasoning/su-01

SU-01 is a research AI model designed to solve math and science olympiad problems at a gold-medal level. Olympiad problems (such as those in the International Mathematical Olympiad or IMO) are among the hardest problems given to humans in mathematics and physics, typically requiring creative multi-step proofs rather than just applying formulas. SU-01 is a 30-billion-parameter sparse model (meaning only about 3 billion parameters are active per computation, which makes it faster to run) trained specifically to produce complete, rigorous, long-form proof-style reasoning. The training recipe has three stages. First, supervised fine-tuning on around 338,000 worked examples, sorted in a specific curriculum order where harder examples come first. Second and third, reinforcement learning in two stages: the first stage rewards finding correct answers, the second rewards the quality of complete proofs and includes a self-refinement step where the model reviews and revises its own work. At inference time, the model can generate extremely long reasoning chains of over 100,000 tokens using a generate-verify-revise loop, which lets it work through hard problems iteratively. With this approach, using test-time scaling (letting the model think longer and vote across multiple attempts), SU-01 reaches 35 points on IMO 2025, which corresponds to gold-medal level. It also reaches gold-medal cutoffs on the USA Mathematical Olympiad 2026 and the International Physics Olympiad. Model weights are available on HuggingFace. The training code uses Docker and the Ray distributed computing framework.