DeepSeek releases world’s first open-source IMO gold-level math AI: DeepSeek-Math-V2

DeepSeek has officially launched DeepSeek-Math-V2, a 685-billion-parameter sparse Mixture-of-Experts (MoE) model that becomes the first open-source large language model to achieve International Mathematical Olympiad (IMO) gold medalist performance. Built on the experimental DeepSeek-V3.2 architecture and released under the Apache 2.0 license, it marks a qualitative leap in AI-driven mathematical reasoning.

The model’s core innovation is a “generate-and-validate” closed-loop system: unlike standard LLMs that output answers in one pass, DeepSeek-Math-V2 integrates a dedicated verifier that scrutinizes every step of its own proofs in real time. If flawed or “lucky-correct” reasoning is detected, the generator self-corrects—mimicking how human mathematicians refine arguments. This self-verification capability is trained end-to-end via reinforcement learning, making reasoning both rigorous and auditable.

Competition Results Highlight Elite Performance:

• IMO 2025: Solved 5 of 6 problems (83.3% accuracy), scoring 210/252—earning a gold medal and ranking 3rd globally, behind only the U.S. and South Korea.
• CMO 2024: Met gold-medal standards.
• Putnam 2024: Scored 118/120 with scaled compute—far surpassing the historical human record of 90.

On Google DeepMind’s IMO-ProofBench, it achieved 99% accuracy on basic problems and 61.9% on hard ones, outperforming all public models and approaching DeepMind’s internal Gemini Deep Think system.

Critically, DeepSeek-Math-V2 is fully open: weights, training details, and inference code are publicly available on Hugging Face and GitHub. Researchers can deploy it locally or in the cloud with one-click support via Transformers.

The training leveraged expert-annotated “pathological proofs” and scaled verification using up to 64 parallel reasoning paths and 16 refinement iterations, enabling high-trust applications in cryptography, formal verification, and drug discovery.

DeepSeek has also released the model’s full solution traces from IMO, CMO, and Putnam, inviting global scrutiny and collaboration.

Links:

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

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