DeepseekMath‑V2 Generates and Verifies Proofs, Aiming to Pop US AI Bubble

In the headline experiments, a single DeepSeekMath‑V2 model is used for both generating proofs and verifying them, with performance coming from the model's ability to critique and refine its own solutions rather than from external math software. For harder problems, the system scales up test‑time compute, sampling and checking many candidate proofs in parallel, to reach high confidence in a final solution. Closing the gap with US labs The release comes on the heels of similar news from OpenAI and Google Deepmind, whose unreleased models also achieved gold-medal status at the IMO, accomplishments once thought to be unreachable for LLMs.

Notably, these models reportedly succeeded through general reasoning abilities rather than targeted optimizations for math competitions. If these advances prove genuine, it suggests language models are approaching a point where they can solve complex, abstract problems, traditionally considered a uniquely human skill. Still, little is known about the specifics of these models.

An OpenAI researcher recently mentioned that an even stronger version of their math model will be released in the coming months. Deepseek's decision to publish technical details stands in stark contrast to the secrecy of OpenAI and Google. While the American giants kept their architecture under wraps, Deepseek is laying its cards on the table, demonstrating that it is keeping pace with the industry's leading labs.

This transparency also doubles as a renewed attack on the Western AI economy, a play Deepseek already executed successfully earlier this year. The strategy seems to be working: As the Economist reports, many US AI startups are now bypassing major US providers in favor of Chinese open-source models to cut costs.