DeepSeek officially released DeepSeek-Math-V2 today. This 685 billion parameter mixture-of-experts (MoE) model has become the world's first open-source mathematical reasoning large model to achieve the level of a gold medalist in the International Mathematical Olympiad (IMO). The model is built upon the experimental architecture of DeepSeek-V3.2 and releases its weights under the Apache 2.0 open-source license, achieving a qualitative leap in mathematical reasoning capabilities.

The most remarkable breakthrough lies in its pioneering "generate-and-validate" closed-loop mechanism. Unlike traditional large models that make a single judgment, DeepSeek-Math-V2 is equipped with a dedicated verifier that performs real-time logical review of each step of the proof generated by the generator. Once a flaw or "lucky correct" faulty reasoning is detected, the verifier immediately provides feedback, and the generator then self-corrects. This process, similar to how human mathematicians refine proofs, is fully internalized into the model's training through reinforcement learning (RL), giving it true "self-verification" capability.

DeepSeek

At the 2025 International Mathematical Olympiad (IMO), DeepSeek-Math-V2 successfully solved 5 out of 6 problems, achieving an 83.3% accuracy rate and earning a gold medal with a score of 210 out of 252, ranking third globally, behind the teams from the United States and South Korea. In the 2024 China Mathematical Olympiad (CMO), the model also reached the gold medal standard. In the Putnam Mathematical Competition, one of the most prestigious undergraduate competitions in North America, it achieved an almost perfect score of 118 out of 120 when testing with unlimited computing power, far surpassing the historical highest human score of 90.

In Google DeepMind's official IMO-ProofBench reasoning benchmark, DeepSeek-Math-V2 achieved a 99% accuracy rate on basic difficulty levels and 61.9% on high difficulty levels, surpassing all previously public models, slightly behind DeepMind's internal Gemini Deep Think enhanced version.

Differing from closed-source systems such as OpenAI's o1 series and DeepMind's AlphaProof, DeepSeek-Math-V2 opens all model weights and complete training details. Researchers and developers can directly download it on Hugging Face and freely deploy it locally or in the cloud. This means that mathematicians and computer scientists around the world can immediately reproduce, audit, or even improve this historic breakthrough.

According to DeepSeek, the training of this model extensively drew on the annotations of "pathological proofs" by human mathematical experts, and then smoothly transitioned from manual to automated through dynamic allocation of verification computing power (up to 64 parallel reasoning paths and 16 iterations). This design not only significantly improved the quality of proofs but also laid the foundation for deploying artificial intelligence in scenarios requiring high trust, such as drug design, cryptography, and formal verification.

Currently, the model is officially available on Hugging Face and GitHub, supporting one-click loading with Transformers. DeepSeek also publicly released the complete problem-solving processes and prediction results of the model in multiple competitions such as IMO, CMO, and Putnam, inviting global peers to examine and verify them.

Address:

https://huggingface.co/deepseek-ai/DeepSeek-Math-V2

https://github.com/deepseek-ai/DeepSeek-Math-V2/blob/main/DeepSeekMath_V2.pdf