DeepSeek has sent shockwaves through the global AI and mathematics communities with the launch of DeepSeekMath V2, an open-weight model that has achieved one of the most remarkable reasoning feats to date. The system successfully solved 5 out of 6 problems at the IMO 2025, placing it on par with the gold-medal-level performance previously demonstrated by Google DeepMind’s Gemini and OpenAI’s experimental reasoning model. In a striking endorsement, Clem Delangue 🤗, CEO of Hugging Face, described the model as “the brain of one of the best mathematicians in the world for free”, underscoring the unprecedented level of access and openness this release represents.
But its capabilities extend far beyond the International Mathematical Olympiad. DeepSeekMath V2 has also delivered exceptional performance on some of the world’s most challenging mathematics assessments. The model posted top-tier results in the China Mathematical Olympiad—widely regarded as one of the toughest national competitions—and nearly achieved a perfect score in the undergraduate-level Putnam exam, earning 118 out of 120, far surpassing the best recorded human score of 90. These achievements highlight the model’s ability not only to generate correct answers but to compete directly with elite human mathematicians across different formats and difficulty levels.
A key pillar of DeepSeek’s approach is its argument that while many AI systems can produce correct outputs, they often lack sound reasoning or the ability to validate their own logic. To overcome this, DeepSeekMath V2 introduces a breakthrough dual-system architecture: a proof generator paired with a dedicated verifier. This verifier rewards the model only when it successfully identifies and corrects its own mistakes, creating a reinforcement loop that steadily improves reasoning quality. By progressively increasing the difficulty of the verification tasks, DeepSeek scales test-time compute and encourages the development of increasingly rigorous mathematical reasoning—a strategic shift away from brute-force problem-solving toward disciplined, self-reflective logic.
The release has been hailed as a milestone for AI accessibility. As Delangue noted, the achievement represents “democratisation of AI and knowledge at its best”, and with the model’s weights already available on Hugging Face, researchers, educators, and developers worldwide can now experiment with and build upon a system that rivals world-class mathematical talent. This launch marks a defining moment in the evolution of open-source AI, bridging elite computational reasoning with global public access.
