DeepSeekMath-V2 Matches OpenAI and Google with IMO Gold Medal Win

DeepSeek has again shattered the exclusive hold of Western tech giants on elite reasoning, releasing an open-weight AI model that matches the performance of OpenAI and Google in mathematics.

Launched Thursday, DeepSeekMath-V2 achieved a Gold Medal standard at the 2025 International Mathematical Olympiad (IMO).

On the William Lowell Putnam Mathematical Competition, the preeminent mathematics competition for undergraduate college students in the United States and Canada, the model scored 118 out of 120, surpassing the top human score of 90. Unlike rival systems hidden behind APIs, DeepSeek has released the weights publicly, allowing researchers to inspect its logic directly.

Arriving during the delay of its flagship R2 model due to US export controls, the release signals technical resilience. It proves specialized architectures can deliver state-of-the-art results even when access to cutting-edge hardware is restricted.

The Gold Standard: Breaking the Proprietary Monopoly

DeepSeekMath-V2 has officially matched the “Gold Medal” standard at the 2025 International Mathematical Olympiad (IMO), successfully solving 5 out of 6 problems. Matching the proprietary benchmarks set by Google DeepMind’s similar milestone and OpenAI’s gold-medal performance, this performance levels the playing field with systems that were previously untouchable.

Far from a simple iterative update, this release represents a fundamental shift in access to elite AI reasoning. While Western laboratories have kept their most capable mathematical models behind “trusted tester” walls or expensive APIs, the model repository for DeepSeekMath-V2 is available for immediate download.

Academic institutions and enterprise researchers can now run the model locally, verifying its capabilities without relying on cloud infrastructure that may be subject to data privacy concerns or geopolitical restrictions.

Beyond the IMO, the model demonstrated unprecedented capability on the Putnam Competition, widely regarded as the most difficult undergraduate mathematics exam in North America. Highlighting the achievement, the DeepSeek Research Team stated:

“On Putnam 2024, the preeminent undergraduate mathematics competition, our model solved 11 of 12 problems completely and the remaining problem with minor errors, scoring 118/120 and surpassing the highest human score of 90.”

Surpassing the human ceiling on such a rigorous exam suggests that the model is not merely retrieving memorized proofs but engaging in novel problem-solving. Achieving 118 out of 120 is particularly notable given the extreme difficulty of the problems, where median scores are historically low.

Independent analysis has further validated these internal metrics. Evaluations on the “Basic” subset of the IMO-ProofBench, a benchmark developed by Google DeepMind, show the model achieving a 99.0% success rate, confirming its reasoning consistency across a broad range of mathematical domains.

Verification is crucial here, as the field has recently been plagued by over-hyped results, such as a retracted claim regarding GPT-5 that falsely alleged the model had solved famous Erdős problems.

By releasing the weights, DeepSeek has effectively commoditized a capability that was considered a major competitive moat for Silicon Valley just months ago. Clement Delangue, Co-founder and CEO of Hugging Face, emphasized the significance of this shift in a post on X:

Under the Hood: The ‘Meta-Verification’ Breakthrough

Historically, the central challenge in mathematical AI has been “hallucination,” where models arrive at the correct answer using flawed, circular, or nonsensical logic. In quantitative reasoning benchmarks, models can often guess the right number without understanding the underlying principles. The DeepSeek Research Team explained the core issue in the technical whitepaper:

“Many mathematical tasks like theorem proving require rigorous step-by-step derivation rather than numerical answers, making final answer rewards inapplicable.”

To address this fundamental limitation, the technical paper details a novel architecture centered on “Meta-Verification.” Unlike standard verification methods that simply check if an answer matches a reference, DeepSeek’s approach evaluates the process of verification itself.

DeepSeek trains a secondary model to judge the quality of the verifier’s analysis, preventing the primary model from “gaming” the reward system by producing convincing-sounding but logically void proofs.

Creating a safeguard against reward hacking, this recursive structure ensures that the model is rewarded only for genuine reasoning rigor. By assessing whether the identified issues in a proof logically justify the score, the system enforces strict logical consistency.

Underpinning this architecture is a “Cold Start” training pipeline. Rather than relying on massive external datasets of formal mathematical proofs, which are scarce and expensive to curate, the model iteratively generates its own training data. Describing the methodology, the researchers state:

“We believe that LLMs can be trained to identify proof issues without reference solutions. Such a verifier would enable an iterative improvement cycle: (1) using verification feedback to optimize proof generation, (2) scaling verification compute to auto-label hard-to-verify new proofs… and (3) using this enhanced verifier to further optimize proof generation.”

“Moreover, a reliable proof verifier enables us to teach proof generators to evaluate proofs as the verifier does. This allows a proof generator to iteratively refine its proofs until it can no longer identify or resolve any issues.”

Through this cycle, the model bootstraps its own capabilities. As the verifier becomes more accurate, it can identify more subtle errors in the generator’s output. Consequently, the generator is forced to produce more rigorous proofs to satisfy the enhanced verifier.

Such dynamics create a positive feedback loop that scales performance without requiring a proportional increase in human-labeled data. At inference time, the model employs “scaled test-time compute.” Instead of generating a single answer, the system generates 64 candidate proofs for a given problem.

It then runs the verification process on all 64 candidates to select the most logically sound path. Shifting the computational burden from the training phase (parameter scaling) to the inference phase (reasoning search), this approach aligns with broader industry trends toward “System 2” thinking where models “ponder” a problem before outputting a solution.

Strategic Resilience: Innovation Despite Sanctions

Serving as a critical counter-narrative to the company’s recent struggles with hardware availability, the release demonstrates significant technical agility. DeepSeek’s flagship R2 model faces hardware-related delays due to persistent failures while training on Huawei’s domestic Ascend chips.

That setback highlighted the immense difficulty Chinese firms face in building a software stack on emerging, unproven hardware under the pressure of US export controls. By pivoting to efficiency-focused architectures, the lab is demonstrating it can still ship state-of-the-art research.

DeepSeekMath-V2 is built on DeepSeek-V3.2-Exp-Base, proving that the sparse attention mechanisms introduced in that model from September are production-ready.

In October, the company launched its optical character recognition tool, which used similar efficiency techniques to compress document processing by tenfold.

Open-weight availability places significant pressure on Western labs to justify their closed-source approach.

As the “moat” of reasoning capability appears to be evaporating, the argument that safety requires keeping these models under lock and key becomes harder to sustain when comparable capabilities are freely available on Hugging Face.

For the broader AI industry, this release suggests that specialized, highly optimized models may offer a viable path forward even when access to massive clusters of Nvidia GPUs is restricted.

By focusing on algorithmic innovations like Meta-Verification and sparse attention, DeepSeek is carving out a competitive niche that relies less on brute-force scale and more on architectural ingenuity.