DeepSeek has unveiled DeepSeekMath V2, an open weight artificial intelligence model that has reached gold medal level performance at the International Mathematical Olympiad 2025. The team behind the model reported that it successfully solved five out of six official problems from the global competition, matching the standards achieved by some of the most advanced closed source systems in the field. The achievement is being described across the AI research community as a significant milestone for open source mathematical reasoning and automated theorem proving.
According to DeepSeek, the model’s capabilities come from a new approach designed specifically for high stakes mathematical reasoning. The model introduces a self verification method where it generates full proofs rather than short answers, and a separate verification module evaluates each step. This mechanism is intended to reduce the chances of flawed reasoning and ensure proof level rigor. The developers said this method helps address the limitations of large language models that often provide confident but incorrect solutions when dealing with abstract mathematical logic.
Along with its Olympiad performance, DeepSeekMath V2 also recorded a score of 118 out of 120 on the 2024 Putnam exam, a competition widely considered one of the most challenging mathematics contests for undergraduates. The reported score places the model above the highest scoring human participants, underlining its capability to tackle advanced symbolic reasoning tasks. The team said the results demonstrate progress toward creating AI systems that can support research level mathematics.
DeepSeek released the model openly, a decision that has been widely welcomed by researchers, educators and developers. Many experts have said that this is the first time an openly available model has matched Olympiad gold standard reasoning while offering full transparency of its internal processes. Several members of the broader AI community have described the release as an important step in democratising high level reasoning tools, as it allows independent researchers to test, adapt and build upon the model’s techniques.
The achievement arrives shortly after reports that other AI systems, including a variant of Google DeepMind’s Gemini Deep Think, also reached gold medal standard on Olympiad problems. However, those models remain closed and are not publicly available for verification or adaptation. DeepSeekMath V2 stands out because its weights, outputs and verification logic are openly shared, providing transparency that is considered essential in scientific and mathematical applications.
Observers note that this development reflects a shifting landscape where open source models are increasingly challenging the dominance of proprietary systems. The availability of a high performing reasoning model can enable participation from universities, independent researchers and smaller labs that may not have access to large scale proprietary systems. Analysts say that this trend could accelerate innovation by distributing advanced research tools more evenly across regions and institutions.
Despite the progress, researchers caution that the model’s performance relies heavily on scaled test time computation. This refers to the large amount of computing power used during the reasoning process to generate and verify proofs. As a result, reproducing the Olympiad level performance may not always be practical for everyday academic or commercial usage. Experts note that the field is still exploring ways to reduce compute requirements without compromising accuracy.
Another important factor is external validation. While self verification improves reliability, independent mathematical review will be necessary to confirm the correctness of the model’s proofs. AI systems working at this level must be checked carefully, especially in fields where accuracy is crucial. Researchers have called for community audits to ensure that the model’s outputs consistently meet the standards expected in formal mathematics.
Even with these caveats, DeepSeekMath V2 is expected to influence future research in fields such as formal verification, cryptography, scientific computing and automated theorem proving. Its ability to generate structured proofs could support mathematicians by helping explore abstract concepts, check long and complex proofs, or act as a tool for learning and training. Educators have noted that openly available reasoning models could provide new pathways for students to engage with advanced problem solving.
Several voices in the global AI research community have described the development as a shift toward more transparent and accessible AI systems. They say that the public availability of such a model could open opportunities for collaboration across academic institutions worldwide, particularly in regions where access to high performance AI systems has been limited.
The milestone also renews debate about the responsibility and governance of advanced reasoning systems. As models become capable of producing research level mathematical outputs, experts highlight the need for oversight, proper documentation and ethical guidelines to ensure responsible use. They say that openness, combined with strong evaluation processes, can help mitigate risks.
DeepSeekMath V2 marks a meaningful step forward for open source AI. By achieving gold medal level performance in one of the world’s most rigorous mathematics competitions and releasing the model publicly, DeepSeek has expanded the possibilities for transparent, collaborative progress in AI driven reasoning. The development signals a potential shift in the field, where open access models begin to play a more central role in advancing scientific and mathematical innovation.