DeepSeek Math-V2 Sets New Standard with Near-Perfect Putnam Score and Gold-Medal IMO Performance

News Summary

Chinese AI startup DeepSeek has released DeepSeekMath-V2, a groundbreaking open-source mathematical reasoning model that has achieved performance levels surpassing many commercial systems. The 685-billion parameter model, built on the DeepSeek-V3.2-Exp-Base architecture, achieved a remarkable 118 out of 120 points on the prestigious Putnam 2024 mathematics competition, exceeding the best human score of 90 points. The model also attained gold-medal-level performance on both the International Mathematical Olympiad (IMO) 2025 and the Chinese Mathematical Olympiad (CMO) 2024.

What sets DeepSeekMath-V2 apart from previous mathematical AI systems is its innovative approach to verification. Rather than simply optimizing for correct final answers, the model employs a sophisticated "verifier-first" architecture that ensures mathematical proofs are not only accurate but also logically rigorous and complete. This represents a fundamental shift in how AI systems approach mathematical reasoning.

The model introduces a novel three-component system: a proof generator that creates mathematical solutions, a verifier that evaluates the quality and soundness of proofs, and a meta-verifier that ensures the verification process itself remains truthful and doesn't hallucinate non-existent errors. This layered approach addresses a critical weakness in previous systems where models could arrive at correct answers through flawed reasoning.

DeepSeek's research team trained the verifier using Group Relative Policy Optimization (GRPO) on over 17,500 proof-style problems from mathematical olympiads and competitions. The system was then enhanced with sequential refinement capabilities, allowing it to iteratively improve proofs across multiple passes within its 128,000-token context window.

On the IMO-ProofBench evaluation developed by Google DeepMind, DeepSeekMath-V2 demonstrated superior performance compared to DeepMind's own DeepThink IMO-Gold system on basic problems and remained competitive on advanced challenges. The model outperformed several leading commercial systems including Gemini 2.5 Pro across multiple mathematical categories including algebra, geometry, number theory, and combinatorics.

Perhaps most significantly for the AI research community, DeepSeekMath-V2 has been released under the permissive Apache 2.0 license, making it freely available for both academic and commercial use. The model can run on systems with 80GB of GPU memory using multi-GPU inference, democratizing access to cutting-edge mathematical AI capabilities.

The competitive landscape reveals interesting dynamics. While OpenAI's GPT-5 maintains an edge in certain benchmarks like the AIME 2025 competition (94% versus DeepSeek's 76%), DeepSeek's open-source model demonstrates that world-class mathematical reasoning capabilities need not be locked behind proprietary systems. Additionally, DeepSeekMath-V2 is substantially more cost-effective, with pricing approximately 40% lower for input tokens and 80% lower for output tokens compared to GPT-5.

The release of DeepSeekMath-V2 represents a significant milestone in the democratization of advanced AI. By achieving gold-medal performance on elite mathematical competitions while remaining open-source and cost-effective, DeepSeek has challenged the assumption that cutting-edge AI capabilities must come from well-funded Western technology giants. The model's success on the Putnam 2024 exam, where it exceeded the best human performance, suggests that AI systems are reaching new levels of mathematical sophistication.

For researchers and developers, the model's availability on Hugging Face with comprehensive documentation and the DeepSeek-V3.2-Exp GitHub repository means immediate practical application is possible. The system's ability to provide not just answers but rigorous, verifiable proofs opens new possibilities for automated theorem proving, mathematics education, and scientific research applications.

The broader implications extend beyond mathematics. DeepSeek's verifier-first approach could influence how AI systems are developed for other domains requiring rigorous reasoning, such as formal verification in software engineering, scientific hypothesis testing, and logical argumentation. The meta-verification concept, which ensures that AI critiques remain honest and grounded, addresses growing concerns about AI reliability and hallucination in high-stakes applications.

Industry observers note that DeepSeekMath-V2's release intensifies competition in the AI sector, particularly as Chinese AI firms continue to produce models that rival or exceed Western counterparts. The model's mixture-of-experts architecture, which activates only 21 billion of its 685 billion parameters during inference, demonstrates sophisticated engineering that balances capability with computational efficiency.

As the AI field continues its rapid evolution, DeepSeekMath-V2 stands as evidence that open-source development can achieve performance levels previously thought to require massive corporate resources. The model's success may accelerate the trend toward open AI development while raising questions about the sustainability of closed-source business models in an increasingly competitive landscape.

For the mathematical AI research community, this release provides a powerful new tool for exploring self-verifiable reasoning systems. The ability to scale test-time compute while maintaining proof quality suggests pathways toward more capable systems that can tackle open mathematical problems without known solutions. Whether this approach will extend successfully to other reasoning domains remains an open and fascinating question for future research.