DeepSeekMath-V2 reaches gold-medal level on IMO 2025 problems

Background

Most math benchmarks used to reward a correct final number or answer. Theorem proving is different: a proof has to be rigorous at every step, and a model can reach a correct conclusion through a flawed argument. DeepSeek's work on mathematical reasoning aimed at checking the reasoning itself, not only the final answer. The model was built on top of DeepSeek-V3.2-Exp-Base ^[1].

What happened

On 27 November 2025 DeepSeek released DeepSeekMath-V2 ^[1][2]. Wikipedia lists the model under the Apache 2.0 license ^[2]. The release was positioned as a self-verifiable reasoning model: one that proves theorems rigorously rather than only producing answers ^[1].

Key details

The central design is a pair of roles. DeepSeek trained an LLM-based verifier for theorem proving and used it as the reward model for a proof generator ^[1]. The generator earns reward by finding and fixing issues in its own proofs before it finalizes them ^[1].

DeepSeek also scaled up verification compute so the verifier could keep pace with the generator. The extra computation was used to label new proofs that are hard to verify, and these labels became training data for the verifier ^[1].

The results were reported on competition problems. DeepSeek said the model reached gold-level scores on IMO 2025 and CMO 2024 ^[1]. On Putnam 2024, a university-level competition, it scored 118 out of 120 with scaled test-time compute ^[1]. Test-time compute is the extra computation spent while the model works on a problem.

Impact

The release showed that a model's own verification can be part of its training loop, with the verifier acting as the reward signal. For readers, the key point is the change of goal: the system is judged on whether its argument holds, not just on whether its final answer matches a key ^[1].