DeepSeekMath-V2 achieves self-verification in Math, matches IMO Gold Medal Standard

DeepSeek has introduced DeepSeekMath-V2, a new mathematical reasoning model that shifts the focus from achieving correct final answers to ensuring rigorous, verifiable reasoning. This model addresses a long-standing challenge in Large Language Models (LLMs); a correct numerical answer does not guarantee the underlying reasoning is sound, a crucial flaw when tackling complex tasks like theorem proving that demand step-by-step logical derivations.

The model is built upon the DeepSeek-V3.2-Exp-Base architecture and is available as an open-source release on Hugging Face. Its core lies in a self-verification framework that allows the AI to critique and correct its own proofs.

Verifier-Generator System drives self-correction

DeepSeekMath-V2 employs a dual-model system, using a proof verifier to assess the rigor and completeness of proofs generated by the proof generator. This structure mimics the self-checking process employed by human mathematicians.

The training incentivizes the generator to identify and resolve as many logical issues as possible in its own proofs before finalizing them. To continuously improve the system, DeepSeek scales the verification compute to automatically label proofs that are difficult to verify. This process creates new training data, ensuring the verifier remains discerning even as the generator becomes more powerful. By prioritizing the rigor of the reasoning process over just the final outcome, DeepSeek aims to build a more faithful and reliable mathematical AI system.

Competition scores rival proprietary models

DeepSeekMath-V2 demonstrates highly competitive performance across major mathematics competitions, often rivaling or surpassing previously set proprietary benchmarks. With scaled test-time compute, the model achieved the following results:

• IMO 2025: Achieved gold-level scores in the International Mathematical Olympiad, putting its performance on par with models from Google DeepMind that achieved similar milestones.

• CMO 2024: Reached a gold-level score on the China Mathematical Olympiad.

• Putnam 2024: Scored a near-perfect 118 out of 120 on the notoriously difficult William Lowell Putnam Mathematical Competition, the undergraduate competition in North America. This score significantly surpasses the highest human score achieved in the competition that year.

On the IMO-ProofBench, a dataset designed to test formal proof capabilities, DeepSeekMath-V2 showed strong dominance on the "Basic" problem set, confirming the consistency of its reasoning.

Open-source availability

DeepSeek released the model weights for DeepSeekMath-V2 under an MIT license, making the technology widely accessible for the research community. This open release challenges the dominance of proprietary systems in this high-stakes area of AI research, allowing researchers to study and build upon a system capable of gold-medal-level mathematical reasoning. Users can download the model from Hugging Face and refer to the DeepSeek-V3.2-Exp GitHub repository for inference support.