Chinese AI startup DeepSeek has just announced the DeepSeek- Math-V2 model, a new step forward in the development of artificial intelligence with the ability to reason toanistically at a high level.
Launched on November 27, local time, the model is specially designed to create and self-verify complex theorems, while demonstrating strong demonstration power thanks to advanced reasoning skills.
DeepSeek- Math-V2 is a deep-end version in the field of toancs, inherited from DeepSeek-V3.2-Exp, a test model that was introduced in September this year.
The entire weight of the model is made public under an open source license Apache 2.0 on major platforms such as hugging Face and GitHub, allowing the research community to access and expand.
According to DeepSeek, the new model is designed around the philosophy of self- control arguments, consisting of two main components, including: a validator that specializes in checking the step of Mathematical proof; a validator capable of self-detecting and self-correcting errors.
This approach helps the model not only create a final solution but also ensure the accuracy of the entire inference process.
In the accompanying technical document, DeepSeek stated that methods based on enhanced learning to optimize the final answer, which used to help LLM achieve high results in competitions such as AIME or HMMT, have now begun to show limitations.
The correct answer at the end does not guarantee that the model is correctly reasoned, especially with proving tasks that require strict reasoning step by step.
Therefore, DeepSeek- Math-V2 is built to expand reasoning capabilities by self-verifying and enhancing logical testing during runtime.
In terms of performance, the Math-V2 has achieved impressive results when tested on the 2025 International Mathematical Olympiad (IMO) and the 2024 CREST Mathematical Olympiad (CMO).
The model is recognized as achieving a score equivalent to the IMO 2025 gold medal, and at the same time achieving 118/120 points in the Putnam 2024 problem, one of the most difficult arithmetic exams in the world.
DeepSeek believes that these results prove that self-versed academic arguments are a feasible development direction, opening up opportunities to build more powerful toanistic AI systems in the future.
Notably, the performance of Math-V2 is assessed to be on par with the models of OpenAI and Google DeepMind, two units that dominate the field of Mathematical AI.
2025 is also the first time IMO has officially recognized the AI models participating, marking the transformation of the Math community.
Although Google was among the first group to be recognized, DeepSeek and OpenAI were not included in the list.
Experts say the development of AI models with in-depth reasoning capabilities can help solve many decades-long problems in fields such as cryptography, theoretical physics or space exploration.
DeepSeek- Math-V2 is expected to be an important stepping stone for the next generation of Mathematical AI.
