DeepSeek AI mathematical reasoning model pioneering self-verifying reasoning

HANGZHOU -- Chinese AI firm DeepSeek has launched DeepSeekMath-V2, a groundbreaking mathematical reasoning model that sets new performance benchmarks and pushes the frontiers of AI-powered problem-solving.

The new model, now open-sourced on Hugging Face and GitHub, introduces a novel self-verifying framework designed to ensure not just correct answers — but logically sound and verifiable proofs.

It demonstrated performances that reached gold-medal levels at both the 2025 International Mathematical Olympiad (IMO) and the 2024 Chinese Mathematical Olympiad (CMO).

Notably, this model also managed to score 118 out of 120 points in the fiercely competitive 2024 Putnam Exam — easily surpassing the top human score of 90.

The model's prowess has been further consolidated via IMO-ProofBench, where it exceeded models like DeepMind's DeepThink.

This system pits two large language models against each other — one acts as a "prover" to generate mathematical proofs, while the other serves as a "reviewer" to scrutinize the reasoning.

Such a mechanism addresses a critical limitation in current AI achievement levels — a correct final answer which does not guarantee a correct reasoning process, according to the DeepSeek team.

DeepSeek said these breakthroughs establish self-verifying math reasoning as a viable and promising path toward developing more powerful and reliable mathematical AI systems.