The Chinese startup DeepSeek has introduced a new AI model, DeepSeekMath-V2, which demonstrates impressive performance in solving complex mathematical problems. The algorithm achieved results equivalent to a gold medal at the International Mathematical Olympiad (IMO 2025) and the Chinese Mathematical Olympiad (CMO 2024).

DeepSeekMath-V2 also scored 118 out of 120 points on Putnam problems, significantly surpassing the best human performance of 90 points. This achievement indicates that language models are approaching the ability to solve complex, abstract tasks traditionally considered uniquely human.

Earlier AI models could provide correct answers but often failed to show the correct reasoning process. DeepSeekMath-V2 solves this issue through a multi-step procedure with a dedicated verifier. This structure allows the algorithm to:

• Check and refine its reasoning steps in real time;

• Criticize and improve its own conclusions without relying on external software (such as calculators or code interpreters).

Thus, the model’s high performance is driven by its reasoning capabilities. For the most challenging tasks, it increases computational resources, generating and verifying many possible proofs in parallel to boost confidence in the final answer.

The release of DeepSeekMath-V2 follows announcements of similar but yet-to-be-released models from OpenAI and Google DeepMind. Notably, the Chinese startup openly shares the technical details of DeepSeekMath-V2 — in sharp contrast to the secrecy maintained by American companies. This transparency shows that DeepSeek is keeping pace with the industry’s leading players.

ORIENT

Also read:

Google DeepMind and OpenAI AI achieve “gold” at the 2025 International Mathematical Olympiad

AI Models Compete in Stock Trading: DeepSeek V3.1 Leads Ahead of GPT and Gemini

DeepSeek and Grok Outperformed GPT-5 and Gemini on the Crypto Exchange, Earning +25%

ChatGPT combines voice and text: How the new update makes AI more natural in conversation