The first open-source model to win a gold medal in the Math Olympiad! DeepSeek's new model has received high praise from netizens: public technical documents, amazing!

DeepSeek's latest open-source mathematical model is pushing towards a stage where it competes alongside tech giants like OpenAI and Google. The DeepSeekMath-V2 model has achieved gold medal level in what is hailed as the world's toughest high school mathematics competition, becoming the first open-source model to accomplish this feat, marking a significant breakthrough for open-source artificial intelligence in complex reasoning capabilities.

Yesterday, DeepSeek announced the launch of its latest mathematical reasoning model, DeepSeekMath-V2, which solved 5 out of 6 problems in the simulated 2025 International Mathematical Olympiad (IMO), achieving gold medal level. This accomplishment makes it the first open-source model to win a gold medal in an IMO-level competition, drawing significant attention from the AI research and developer community.

This performance directly competes with industry giants. Just this July, Google DeepMind's advanced version of Gemini and an experimental reasoning model from OpenAI also reached the gold standard for IMO 2025, solving 5 problems as well, making them among the first AI models to achieve this level. However, unlike the closed-source experimental models from Google and OpenAI, the model weights of DeepSeekMath-V2 are publicly released under the Apache 2.0 license and are available for public download.

It is worth mentioning that DeepSeekMath-V2 employs an innovative self-verification training framework. The core of this method is to train a dedicated "verifier," whose task is to assess the quality of the proof process rather than the correctness of the final answer. Moreover, to prevent the model from overfitting its own checking mechanism, DeepSeek continuously increases the difficulty of the verification process by adding computational load and automatically tagging hard-to-verify proofs, ensuring that the verifier and generator evolve in sync.

This move is seen as an important step towards the democratization of artificial intelligence. The release of this model not only proves that the open-source community is capable of catching up with or even matching top closed-source laboratories in cutting-edge AI research but may also reignite discussions in the market about whether open-source models will erode the commercial moats of closed-source products—a topic that once shook investors' confidence in AI giants like Nvidia.

Joining the Ranks of the Elite: Competing with OpenAI and Google

The outstanding performance of DeepSeekMath-V2 signifies that it stands on equal footing with the world's leading AI laboratories in the complex field of mathematical reasoning. The International Mathematical Olympiad (IMO) is generally regarded as the most challenging high school mathematics competition globally, with only 72 out of 630 human participants achieving gold medals in the 2025 competition.

In addition to its achievements in IMO 2025, the model has also demonstrated top-level performance in other high-difficulty mathematics competitions. According to DeepSeek, it has also reached gold medal level in China's top national competition—the China Mathematical Olympiad (CMO).

In the Putnam Mathematical Competition (Putnam 2024), aimed at undergraduate students, the model completely solved 11 out of 12 problems, with only minor errors in another problem, ultimately scoring 118/120, surpassing the highest human participant record of 90 points

Milestone of Open Source: Community Praises "Remarkable Release"

Compared to the experimental models that Google and OpenAI have yet to publicly disclose, the core appeal of DeepSeekMath-V2 lies in its complete openness. The model's weights have been released on the open-source community platform Hugging Face, allowing researchers and developers to download them freely.

Clement Delangue, co-founder and CEO of Hugging Face, praised on social platform X: "Imagine being able to freely access the brain of one of the world's greatest mathematicians."

He added, "As far as I know, no chatbot or API has previously allowed you to access a model at the IMO 2025 gold medal level." He emphasized that users can explore, fine-tune, and optimize the model without restrictions, running it on their own hardware, "with no company or government able to take it back. This is the best embodiment of the democratization of artificial intelligence and knowledge."

Another user, elie, also commented: "Is DeepSeek Math V2 the first open-source model to achieve gold medal level at the IMO? And we also got a technical report; this is truly a remarkable release."

Other users commented that they liked 5-7 ideas, each relatively simple, stacked continuously, resulting in unexpectedly better outcomes, looking more like engineering than research.

Self-Verification Framework: Beyond Answers, Focus on the Reasoning Process

DeepSeek pointed out in its technical report that recent artificial intelligence models, while adept at obtaining correct answers in mathematical benchmark tests, often lack rigorous reasoning processes. The report stated: "Many mathematical tasks, such as theorem proving, require rigorous step-by-step derivation rather than just a numerical answer."

To address this issue, DeepSeekMath-V2 employs an innovative self-verification training framework. The core of this method is training a dedicated "verifier," whose task is to assess the quality of the proof process rather than the correctness of the final answer. Subsequently, this verifier is used as a reward model to guide an independent "proof generator." The generator only receives rewards when it successfully identifies and corrects errors in its own proofs This mechanism incentive model aims to discover and resolve as many issues in its reasoning chain as possible before arriving at a final answer. DeepSeek emphasizes that "self-verification is particularly important in test-time compute for open-ended problems without known solutions." Test-time compute refers to allocating substantial computational resources during the reasoning phase, allowing the model more time to reason, explore multiple solutions, and refine its answers.

Dynamic Evolution System: Solving the "Self-Overfitting" Dilemma

To prevent the model from overfitting its own checking mechanism—i.e., only learning to deceive its own validator—DeepSeek employs a dynamic evolutionary strategy. The team continuously increases the computational load and automatically labels difficult-to-verify proofs to enhance the difficulty of the verification process, ensuring that the validator evolves in sync with the generator.

DeepSeek explains in its technical documentation that this approach allows them to "scale verification compute to automatically label new, difficult-to-verify proofs, thereby creating new training data to further improve the validator." Through this verification-generation closed loop and meta-verification mechanism, the model can achieve fully automated data labeling and continuous performance optimization, validating the feasibility of self-driven learning systems in solving complex mathematical reasoning tasks