Deepseek is pushing deeper into advanced mathematical reasoning with DeepseekMath-V2, a model the Chinese startup says has reached gold medal-level results in major math competitions. The claim places Deepseek in the same conversation as Western AI labs that have recently reported similar progress with unreleased systems.
The bigger issue is not only whether an AI can land on the right answer. Deepseek is presenting DeepseekMath-V2 as a system designed to check the quality of its reasoning, especially in proof-heavy problems where a correct final result is not enough.
What Deepseek says the model achieved
According to Deepseek, DeepseekMath-V2 reached gold medal-level results at the International Mathematical Olympiad (IMO) 2025 and the Chinese CMO 2024. In the Putnam competition, the model scored 118 out of 120 points, above the best human result of 90 points.
Those benchmarks matter because they test more than routine calculation. Math competitions such as the IMO and Putnam reward precise reasoning, proof construction, and the ability to handle abstract problems under strict standards.
The source does not describe the model as using calculators, code interpreters, or other external tools for these headline results. It says the paper never mentions such tools, and that the setup suggests the benchmark performance comes from natural language alone.
That distinction is important. A model that depends on external software to compute answers is different from a model that can reason through a proof in language, identify weaknesses, and refine its own solution.
Why proof verification is central
Deepseek’s technical documentation frames the problem directly: previous AIs could sometimes produce a correct final answer while failing to show valid work. In mathematics, that is a serious limitation, because the path to the answer is part of the answer.
DeepseekMath-V2 uses a multi-stage process to address that issue. A "verifier" evaluates the proof. A "meta-verifier" then checks whether the criticism from the verifier is justified.
This creates a loop in which the model can generate, inspect, challenge, and improve solutions. In the headline experiments, a single DeepSeekMath‑V2 model is used both to produce proofs and to verify them.
The performance therefore comes from the model’s ability to critique and refine its own output, rather than from external math software. For difficult problems, the system also increases test-time compute by sampling and checking many candidate proofs in parallel before settling on a final solution with high confidence.
In plain language, Deepseek is trying to make the model behave less like a fast answer generator and more like a careful problem solver. The model is not simply asked to respond; it is pushed to examine whether its own reasoning holds up.
How this fits into the AI race
The release follows similar news from OpenAI and Google Deepmind, whose unreleased models also reached gold-medal status at the IMO. Those results were once thought to be unreachable for LLMs.
The source says those models reportedly succeeded through general reasoning abilities rather than targeted optimizations for math competitions. If these advances are genuine, they suggest language models are moving closer to solving complex, abstract problems that have traditionally been considered a uniquely human skill.
Still, there is a major difference in what the companies are sharing. Little is known about the specifics of the OpenAI and Google models. An OpenAI researcher recently mentioned that an even stronger version of their math model will be released in the coming months.
Deepseek, by contrast, has published technical details. That openness gives the release a second purpose: it is not only a research claim, but also a signal that Deepseek wants to be seen as keeping pace with the industry’s leading labs.
Why openness changes the business stakes
Deepseek’s transparency also lands inside a broader economic contest. The source describes it as a renewed attack on the Western AI economy, a strategy Deepseek had already executed successfully earlier this year.
The pressure appears to be showing up in customer behavior. As the Economist reports, many US AI startups are now bypassing major US providers in favor of Chinese open-source models to cut costs.
That dynamic makes DeepseekMath-V2 more than a technical release. If a Chinese startup can publish strong technical details while reporting frontier-level math performance, it challenges the idea that the most capable AI systems must remain closed, expensive, and concentrated among a small group of US providers.
At the same time, the rivalry cuts both ways. As models become more capable, their development becomes more politically charged. The source argues that this shift could strengthen US labs, because Deepseek’s push at the frontier may help OpenAI and its peers justify the speed and scale of their own advances.
What to watch next
The main question now is how much confidence researchers and users can place in reported math benchmarks across competing labs. Deepseek has put more technical information on the table than some rivals, but the broader field still has limited visibility into the strongest systems from OpenAI and Google Deepmind.
For now, DeepseekMath-V2 stands out for three linked reasons:
- Deepseek reports gold medal-level results at the International Mathematical Olympiad (IMO) 2025 and the Chinese CMO 2024.
- The model scored 118 out of 120 points in the Putnam competition, compared with the best human result of 90 points.
- Its method emphasizes proof generation, verification, and self-correction rather than external math tools.
That combination points to a wider shift in AI reasoning. The competition is no longer only about producing fluent answers. It is increasingly about whether models can build arguments, test their own work, and make abstract reasoning reliable enough to trust.