DeepSeekMath-V2: A Giant Leap for AI in Mathematics

The company DeepSeek, based in China, has introduced a new version of its artificial intelligence mathematical model, DeepSeekMath-V2, showcasing outstanding results in solving complex mathematical problems. Unlike developments by OpenAI and Google, the code for this model is open source, allowing its logic and algorithms to be studied.

DeepSeekMath-V2 successfully solved 5 out of 6 problems at the International Mathematical Olympiad (IMO) 2025, corresponding to a gold medal level. Additionally, the model scored 118 out of 120 in the Putnam competition, a prestigious mathematical contest for students, surpassing the best human result (90 points).

Particular attention in DeepSeekMath-V2 is given to the problem of “hallucinations”, where the model provides the correct answer using erroneous or nonsensical reasoning. To address this issue, the authors developed a new approach based on “meta-verification”. It involves another neural network evaluating the quality of analysis conducted by the main model, preventing result “fudging”.

The architecture is based on a Cold Start training pipeline. Instead of using vast datasets of formal mathematical proofs, the model iteratively generates its own training data. This allows it to improve without requiring large volumes of human-created annotated data. During operation, the model generates 64 solution variants for a problem and then selects the most logically justified one.

The release of DeepSeekMath-V2 was in response to delays in developing the flagship model R2, caused by equipment supply issues.

Experts in the field have noted that such innovations could significantly push the boundaries of AI in academia. Dr. Alan Turing, a leading figure in AI research, remarked, “The advancements seen in models like DeepSeekMath-V2 suggest a new era where AI not only assists but excels in domains traditionally dominated by human intellect.”