A 4-billion-parameter model now solves Olympiad math problems cheaper than the giant proprietary systems

Researchers built a tiny open-source AI model that can prove difficult mathematical theorems at the level of International Mathematical Olympiad problems, matching or beating much larger closed systems at a fraction of the computational cost. This means the expensive, secret training methods behind proprietary math AI aren't actually necessary — smaller, reproducible models can reach the same performance.

For years, the only systems that could solve hard mathematical problems were proprietary black boxes running on massive hardware, making it impossible for researchers to understand how they worked or improve them. This paper shows that's not true. A tiny open model trained with straightforward techniques (supervised learning, reinforcement learning, a reasoning cache) can match those systems. That matters because it breaks the assumption that frontier math AI requires proprietary scale and secrecy. Now researchers outside the largest companies can study and modify the training pipeline itself, not just call an API.