Breaking News: DeepSeek Open-Sources New Model and Claims Victory in Math Olympiad Proof Challenge

A self-verifying AI system is now one step closer to tackling research-level mathematics.

On November 27, DeepSeek officially open-sourced its new model DeepSeekMath-V2, an “Olympiad gold-medal-level” AI system with exceptional theorem-proving abilities.

DeepSeekMath-V2 delivered gold-standard performance in both the 2025 International Mathematical Olympiad (IMO 2025) and the 2024 Chinese Mathematical Olympiad (CMO 2024). Even more impressively, it earned an almost perfect score of 118/120 in the 2024 Putnam Mathematical Competition, dramatically surpassing the highest known human score of 90.

As shown in performance benchmarks, DeepSeekMath-V2 surpassed Google’s IMO-winning model DeepThink by a decisive 10% margin on key proof benchmarks such as IMO-ProofBench.

These results reveal something groundbreaking: self-verifiable mathematical reasoning is not only feasible, but may be the key to building dramatically stronger AI-driven mathematics systems.

Following DeepSeek’s usual practice, the model was deployed immediately after release. Upon initial testing, DeepSeekMath-V2 rapidly delivered correct proofs for classic problems such as proving √2 is irrational. When asked, “Are there more odd numbers or even numbers?”, it provided a correct and structured proof that even non-experts can follow. Of course, Olympiad-level proofs are far more complex, and mathematically advanced readers are encouraged to explore and experiment with the model themselves.

Turning to the research behind the model, traditional RL-based approaches can achieve high performance on competitions that evaluate only final answers, such as AIME or HMMT. However, DeepSeek identifies two critical limitations of this approach:

First, final answers alone do not reliably indicate reasoning correctness. A model may reach the correct answer by flawed logic or lucky manipulation.
Second, traditional RL methods fail for theorem proof tasks, where the derivation process — not merely the final answer — determines the validity.

To solve this, DeepSeek focused on developing explicit proof-verification skills within large language models. Built upon DeepSeek-V3.2-Exp-Base, DeepSeekMath-V2 is trained to understand its reward mechanisms and to pursue correct reasoning intentionally, rather than by blind trial-and-error.

DeepSeek established an advanced evaluation framework for mathematical proofs, enabling the model to mimic expert-style judgment. Using a version of DeepSeek-V3.2-Exp-SFT as a base, the model was trained via reinforcement learning with two core reward components: a formatting reward and a scoring reward.

The training dataset included 17,503 competition problems, AI-generated proof candidates, and expert-scored proof samples — forming a large-scale RL training corpus. DeepSeek then defined RL objectives for both the generator and the verifier models. While this enabled strong analytical evaluation, the researchers discovered a potential vulnerability: the verifier might “game the system” by producing correct-looking scores while fabricating nonexistent reasoning issues.

To counter this flaw, DeepSeek implemented meta-verification, a second-layer analytical evaluation that significantly increases the fidelity of proof assessment.

During proof generation, DeepSeek trained the model to reason and self-check progressively rather than declaring correctness unconditionally. This ensured that the model does not simply insist that its proof is valid even when the external verifier flags errors.

Ultimately, DeepSeek’s verifier and generator form a cooperative feedback loop: the verifier improves the generator, and the generator continually produces new proof attempts that challenge the verifier. These challenges become valuable training data, pushing both components forward. In simple terms: the verifier analyzes and checks each reasoning step, while the generator learns from its mistakes and refines itself.

Experimental results show that DeepSeekMath-V2 consistently outperformed GPT-5-Thinking-High and Gemini 2.5-Pro across all CNML mathematical categories, including algebra, geometry, number theory, combinatorics, and inequalities.

In sequential optimization with self-verification, DeepSeekMath-V2 substantially improved proofs on 2024 IMO candidate problems, demonstrating that the model is capable of reliably distinguishing strong proofs from flawed ones and refining its reasoning autonomously.

In high-compute exploration, the system solved 5 out of 6 IMO 2025 problems, and 4 full plus 1 partial solution from CMO 2024 — achieving gold-medal-level results, outperforming DeepMind’s DeepThink on foundational sets while remaining highly competitive on advanced sets.

Nonetheless, DeepSeek observes that the most difficult IMO-tier problems remain a challenge — even for this advanced model.

Importantly, on unsolved problems, DeepSeekMath-V2 frequently identifies the genuine reasoning obstacles within its own proofs, while fully solved problems pass all 64 verification rounds without contradiction. This demonstrates that the trained verifier is capable of assessing proofs previously considered extremely difficult to evaluate automatically.

With increased computational depth during inference, DeepSeek’s model can solve problems that would typically require human contestants several hours of dedicated reasoning.

Conclusion: A self-verifying AI architecture may be the key to solving research-grade mathematics.

In summary, DeepSeek has built a model capable of both generating and validating mathematical proofs. The team has moved beyond simplistic final-answer reward systems, toward a sophisticated framework of self-evaluating mathematical reasoning.

This work shows that large language models can develop meaningful meta-cognitive abilities for complex reasoning tasks. Although challenges remain, this research direction brings AI significantly closer to the long-term goal: self-verifying AI systems capable of solving true research-level mathematical problems.